Fashion retrieval is the challenging task of finding an exact match for fashion items contained within an image. Difficulties arise from the fine-grained nature of clothing items, very large intra-class and inter-class variance. Additionally, query and source images for the task usually come from different domains - street photos and catalogue photos respectively.
Due to these differences, a significant gap in quality, lighting, contrast, background clutter and item presentation exists between domains. As a result, fashion retrieval is an active field of research both in academia and the industry. Inspired by recent advancements in Person Re-Identification research, we adapt leading ReID models to be used in fashion retrieval tasks.
We introduce a simple baseline model for fashion retrieval, significantly outperforming previous state-of-the-art results despite a much simpler architecture. We conduct in-depth experiments on Street2Shop and DeepFashion datasets and validate our results. Finally, we propose a cross-domain (cross-dataset) evaluation method to test the robustness of fashion retrieval models.
Fashion retrieval is the challenging task of finding an ex- act match for fashion items contained within an image. Difficulties arise from the fine-grained nature of clothing items, very large intra-class and inter-class variance. Additionally, query and source images for the task usually come from different domains - street photos and catalogue photos respectively. Due to these differences, a significant gap in quality, light- ing, contrast, background clutter and item presentation exists between domains. As a result, fashion retrieval is an active field of research both in academia and the industry.
Inspired by recent advancements in Person Re-Identification research, we adapt leading ReID models to be used in fashion retrieval tasks. We introduce a simple baseline model for fashion retrieval, significantly out- performing previous state-of-the-art results despite a much simpler archi- tecture. We conduct in-depth experiments on Street2Shop and DeepFash- ion datasets and validate our results. Finally, we propose a cross-domain (cross-dataset) evaluation method to test the robustness of fashion re- trieval models.
Fashion image retrieval, commonly associated with visual search, has received a lot of attention recently. This is an interesting challenge from both a commercial and academic perspective. The task of fashion retrieval is to find an exact or very similar products from the vendor’s catalogue (gallery) to the given query image. Creating a model that can find similarities between content of the images is essential for the two basic visual-related products for the fashion industry: recommendations and search.
Visual recommendations (VR) - visual similarity needs to be found between a currently viewed product and other products in the database. Retrieval is done among the images from a single domain, which is a domain of catalogue photos.
Visual search (VS) - visual similarity between user taken/uploaded photo and the products’ photos in the vendor’s database. This case is a cross-domain retrieval as the photos from catalogue contains generally a single item and are taken by professionals using high-quality equipment, assuring proper background (usually solid white) and lighting conditions. On the other hand, photos taken by users are taken with a smartphone camera in uncontrolled lighting conditions and are of lower quality, noisy, with multiple persons and garments.
From a business perspective, especially for e-commerce retailers, these tools still have a lot of untapped potential. VS in particular can be used in various ways to enrich the customer experience and facilitate a more convenient search, since it may be easier and more natural for users to take a photo of a product and upload it directly via an app rather than to use textual search.
During work on VS solutions, we found that the task of visual search for clothes is in many ways analogous to Person Re-Identification problem (ReID). Yet we have not encountered any work that examines models specific to Person ReID tasks in the context of fashion image retrieval tasks. Thus, we decided to modify and apply approaches from Person ReID field to fashion retrieval.
There are four main differences between the ReID and VS problems. Firstly, in ReID tasks data may be regarded as homogeneous, since the query and gallery images are both taken by CCTV cameras, therefore they are from the same do- main. The cameras, resolution, view angles, and other factors may vary within the domain. In fashion visual search tasks images come from two domains. One domain contains store catalogue photos taken by professionals at the same studio settings. The second domain consists of photos taken by users, which may vary due to lighting conditions, the quality of the camera, angles and focal length. Additionally, the content of the image may be photographed differently. Cat- alogue photos are often well aligned, the background is usually solid white or monochromatic and the fashion item is fully visible. In contrast, user images often contain multiple objects, a cluttered background and the fashion item is often occluded or cropped.
Secondly, ReID problems may use additional temporal information to narrow down the search space, e.g., which adds an information stream unavailable in the fashion domain.
Moreover, fashion items, especially clothes, are highly deformable goods re- sulting in a high intra-class variance in appearance, further complicating appear- ance comparisons.
Finally, in ReID, the majority of images contain a whole person and can be relatively easily aligned. User images in fashion may differ both in the alignment, orientation and content. They may contain a whole person with multiple clothing items, upper or lower half of a person, clothes laying on the sofa etc.
We also pinpointed four major similarities between ReID and clothes re- trieval. In both domains, the core problem is an instance retrieval task that relies heavily on fine-grain details, which makes both tasks complex and challenging. Secondly, both domains must implicitly consider a common subset of clothes & materials deformations specific to human body poses and shapes. Moreover, in both domains the models have to deal with occlusion, lighting, angle and ori- entation differences. Finally, clothes are also an important factor in ReID task, especially when faces are not visible.
Due to the similarities and despite the differences of the two aforementioned tasks we decided to investigate how and which ReID solutions can be effectively applied to the fashion retrieval task. Our main contributions are:
- Recognition of deep analogies between Person ReID and Fashion Retrieval tasks.
- A review of ReID models and their applicability to Fashion Retrieval tasks.
- Adjustments of ReID models for their application to the fashion domain.
- Thorough evaluation of the adapted ReID models on two commonly used fashion datasets.
- A simple baseline model for fashion retrieval significantly outperforming state-of-the-art results for DeepFashion and Street2Shop datasets.
- A cross-domain (cross-dataset) evaluation method to test robustness of the solutions.
In the following sections we provide a review of Person Re-Identification models and solutions used in Fashion Retrieval domain. We show the baselines for our comparisons (Section 2). Then, we provide an argument for our selection of applicable ReID models (Section 3). Subsequently, we describe our results, compare them with previous state-of-the-art approaches and provide a discussion (Section 4) of our findings. We conclude with our best achievements (Section 5). In supplement material we provide statistics of datasets used for training and evaluation, as well as additional samples from model outputs.
2 Related Work
This section focuses mainly on the problems of extracting features from images and the retrieval process itself. Instance retrieval is basically a two-stage process: (1) encoding image/item into a n-dimensional embedding vector; (2) searching for the most similar product embedding to the query in the n-dimensional space under a chosen distance metric.
2.1 Feature Extraction
The process of extracting features from an image may be defined as passing the pixel matrix through a function to obtain an embedding vector that encompasses all necessary information about the image in the encoded form. It is the first and most indispensable step towards image retrieval. Most research is devoted to improving the performance of this stage.
Over the years, the functions used for encoding image content have evolved. At first, hand-crafted features describing global image characteristics (e.g. his- togram of oriented gradients - HOG) or local dependencies like SIFT (scale- invariant feature transform) were used. These methods did work, however sub- sequently convolutional neural networks (CNNs) surpassed these methods in computer vision tasks and dominated the field.
One of the first works using CNN for an image retrieval task was  that also managed to surpass the stat-of-the-art of approaches based on SIFT, HOG and other hand-crafted features. In  authors used CNNs extracted features and trained a model using a triplet loss function, which was widespread by  for face recognition task. By ’CNN extracted features’ we mean the final output of either one or many convolutional layers that are arranged in various combinations (e.g. pyramid, sequence of layers). All of these combinations work as various learned filters for an image matrix.
2.2 Image Retrieval
Image retrieval may be defined as a task of matching images based on their feature vectors/embeddings under a chosen distance metric. There are three main approaches when dealing with image retrieval using deep learning embeddings: direct representation, classification loss training and deep metric learning. We explain them in the following paragraphs.
Direct representation This is the simplest approach as it does not involve any training on the target dataset. In most cases, an ImageNet pre-trained neural network is used. To obtain the feature vector, an image is fed into the network and the output from one or more layers is treated as its embedding. To find the most similar images to the query, a cosine or other similarity metric is applied to the set of embeddings. Such an approach was used in .
Classification loss training Another approach is to again take a pre-trained CNN and train it on the target dataset using a classification loss function. The dif- ference to the direct representation is that the network is further trained (fine- tuned) on a target dataset. Thus, the embeddings should be optimized in a way to allow correct classification and in turn indirectly improve the retrieval quality.
This approach was also present in previously mentioned reference , where the authors used pre-trained AlexNet to learn the similarity between street and shop photos using cross-entropy loss over pair of images, classifying them as either matching or non-matching.
Deep metric learning Image retrieval formulated as a classification problem is not an optimal approach, as the classification is done mainly in the last fully connected (FC) layer of the neural network (NN), which is not used during the inference stage. Moreover, the gradient flow does not explicitly optimize the feature embeddings to be similar/dissimilar under the distance metric used for retrieval.
Deep metric learning treats the image retrieval problem somewhat as a clus- tering or ranking problem and optimizes the mutual arrangement of the em- beddings in space. One of the most widely used approaches that pulls the same class closely and pushes away other class embeddings is using a triplet loss for training neural network. The triplet loss and its’ modifications are described in our supplement in more details.
2.3 Fashion Image Understanding
Fashion image understanding has been an area of research for a while. The papers that introduced two modern large public datasets , , gave the start for a rapid development of machine learning applications in the fashion field. Over the years numerous fashion datasets were released by various authors: , , , , , ,  accelerating the research and allowing a wider range of tasks to be tackled.
Clothes retrieval seems to be the most popular task , , , , . However, it is often combined with another task to attain a synergy effect, either with landmark detection (detecting keypoints of clothes) , attribute predic- tion , , or object detection , , , .
The problem of Person Re-Identification, being an instance retrieval task, has undergone a similar evolution as the image retrieval domain, starting from using hand-crafted features  and matching their embeddings under chosen distance metric. Currently CNN extracted features are the dominant approach.
Similarly to Section 2.2 a variety of loss functions are used in ReID prob- lems: classification loss , verification loss  and triplet loss . Triplet Loss performs best  in most cases by a significant margin.
Due to the fact that the ReID problem focuses on images of people from CCTV cameras, who are mostly in a standing position, some works exploited this fact by adding this information during training. In  the authors used horizontal pooling of image stripes, while  modified a network architecture search algorithm specifically for ReID task, that also integrated the human body in an upright position during training. ,  proposed using human/skeleton pose information along with global and local features for better alignment of features, thus, increasing retrieval performance.
Moreover, Person Re-Identification problem settings allow to implicitly or explicitly use spatio-temporal information, as time constraint allows to eliminate a portion of irrelevant images during retrieval stage , . Currently spatio- temporal approach  tops the leaderboard in Person ReID task on Market1501 - common dataset for Person Re-Identification task.
3 Our Approach
The aim of our work was to investigate if and how ReID models can be success- fully applied to a fashion retrieval task. Based on the performance of various models in Person Re-Identification task and their suitability for the fashion re- trieval problem, we selected the most appropriate models. The motivation behind our final choice was threefold:
- The chosen models should exhibit top performance on ReID datasets.
- The chosen models should cover different approaches with regards to the network architecture, training regime etc.
- The code had to be publicly available.
3.1 ReID Models
Two models were chosen based on the above criteria.
First,  presents an approach that combines the most efficient training
approaches and sets a strong baseline for other ReID researchers, hence its name: ReID Strong Baseline (RST). Through a thorough evaluation it showed that by combining simple architecture, global features and using training tricks (warm- up learning rate, random erasing augmentation, label smoothing, last stride, BNNeck, center loss; for more details see ) one can surpass state-of-the-art performance in the ReID domain.
The second choice -OSNet-AIN  uses features extracted from different scales, which seems to be a well-suited approach in a fine-grained instance re- trieval. The authors devised a flexible fusion mechanism that allows the network to focus on local or global features that improves performance. This is a different approach from the first one, as both global and local features are used and the whole neural architecture was purpose-built for ReID task.
It is worth mentioning that some top scoring papers from website Paper- swithcode were not chosen as they either required temporal data  or made an implicit assumptions of spatial alignment of objects in images . Neither approach was suitable to the clothes retrieval task and the datasets we used for evaluation.
3.2 Settings for Our Experiments
To fully explore ReID models and their application to the fashion retrieval task we conducted a number of experiments with the two selected approaches. We ex- plored several aspects with potential to influence the performance in the fashion domain:
- backbone architecture (a feature extractor part of a model; see Figure 2)
- triplet and quadruplet loss functions in RST
- size of input images
- re-ranking (post-processing methods to improve accuracy of retrieval)
- cross-domain evaluation (training on one dataset and testing on another)
Using the selected ReID models we dealt with three loss functions. As for the OSNet-AIN, we followed the training regime proposed by its authors and we trained the model using only a classification loss. For the RST model we used a loss function that consisted of three elements: a classification loss, a triplet loss and a center loss. Additionally, we tested if by replacing the triplet loss with a quadruplet loss one cloud improve the performance.
3.3 SOTA Models for Clothes Retrieval vs RST Model
A detect-then-retrieve model for multi-domain fashion item retrieval  showed SOTA performance for Street2Shop dataset. The model consists of Object De- tector (Mask-RCNN), which is used to crop clothing items from shop images and a Retrieval Model that is built upon Resnet-50 backbone, but incorporates RMAC pipeline . The RMAC is a global representation of images that uses CNN as a local feature extractor. Local features are extracted and max-pooled from multiple multi-scale overlapping regions covering whole image, creating a single feature vector for each region. These region features are aggregated, nor- malized, transformed and then normalized again. Moreover, in  an ensemble model was also created that consisted of two single models, one trained using triplet loss and other one using AP loss. Input image sizes used in  were also much larger than used in our RST model as the images were resized to 800 pix- els, which caused problem in processing them thought the network, since they were forced to process either a single image (AP loss) or a single triplet (triplet loss) at a time. For more details see .
Fashion Retrieval via Graph Reasoning Networks on a Similarity Pyramid  achieved SOTA performance on DeepFashion dataset. They based their solution on graphs and graph convolution. The architecture consists of three parts. First is a CNN network, which extracts multi-scale (pyramid) features from numerous, overlapping windows that cover the whole image. What is important two images of the same item are fed into the network street and shop images and they are processed together as a pair. Second part is the Similarity Computation, which computes region similarity between street and shop images for all possible local feature combinations at the same pyramid scale. Next, a graph is built. Com- puted region similarities are the nodes, the relations between region similarities are the edges. Scalar edges weights are computed automatically based on the node, incoming and outgoing edges. Finally, the convolution operations are ap- plied to the graph and classification loss is used for training. The aim of the network is to classify the pair of images as depicting the same clothes or not. Input size of images was 224x224, which is slightly smaller than used in our best performing RST model. For more details see .
RST model Compared to the two models described above, the RST model can be characterised by its simplicity. Even though, the architecture is much simpler, the performance exceeds significantly current SOTA. The RST model consists of 3 parts. First, CNN backbone extracts features from images and global average pooling is applied to create global feature vectors. They are used for computing quadruplet and center loss. Next, the global feature vectors are normalized and we call them Images embeddings (see Figure 2). Images embeddings are used during training as input to a fully connected layer, while during inference stage (retrieval) they are used to compute similarity distance. In Figure 2 we presents the RST model, which shows the pipeline and all parts of the architecture. It can be deemed as strikingly simple, yet it substantially exceeds current SOTA results.
In this section we describe our methodology and details of experiments we ran. We conducted evaluation of our approach and its settings on Street2Shop and DeepFashion datasets.
For both Street2Shop and DeepFashion we cropped the clothing items of interest from the whole images, as both datasets contain ground-truth bounding boxes for street photos. Additionally DeepFashion contains bounding boxes for shop images, so we cropped them as well. Street2Shop does not provide bounding boxes for shop images, therefore we used the original, un-cropped shop images for the gallery set. These settings for street images are in line with those used in  and , which we consider as the current SOTA for Street2Shop and DeepFashion respectively.
To adapt the RST model to fashion datasets we introduced some changes to the code, which were mostly required due to significant difference in size between ReID and clothes retrieval datasets. The size of DeepFashion and Street2Shop caused RAM and video RAM overflow, which was solved by using more memory- efficient methods. Moreover, we added some new functionalities and scripts. Our improvements are described in detail in the supplement.
Unfortunately, evaluation with re-ranking and without category constraint for Street2Shop was impossible despite improving the code efficiency, therefore we decided to conservatively estimate the results. The estimated results are marked with asterisk (*). For more details see the supplement.
4.2 ReID Models Comparison
In the first experiment we trained both ReID models to compare their perfor- mance on DeepFashion and Street2Shop datasets. For training we used 256x128 input images. Table 1 contains the results of the models and current state-of-the- art approach. It can be seen that the RST model surpasses the current SOTA model and the OSNet-AIN performs worse than the other two approaches. The reason for poor performance of OSNet may be the fact that it is a lightweight network and built very specifically for ReID tasks.
Due to the large performance gap between RST and OSNet models, we de- cided to conduct further experiments using only the more promising RST model.
4.3 Backbone Influence
In this section we inspect the influence of the used backbone on the results. The results of our experiments are shown in Table 2. All runs during this experiment were performed on input images of size 256x128, using an RST approach with all tricks enabled and using quadruplet loss. All models were trained for 120 epochs on 1 GPU, with base learning rate 0.0001 and batch size 128. SeResNext-50 , EfficientNet-b2 and EfficientNet-b4  were trained on 2 GPUs using ModelParallel mode due to their size.
Our findings are in line with backbone performance presented by the authors in , i.e. ResNet50-IBN-A  is the best performing backbone. Regarding the results, we infer that such a large advantage in performance for Resnet50-IBNs may be caused by instance normalization, which reduces the impact of variations in contrast and lighting between street and shop photos.
4.4 Influence of Loss Functions
The results for the RST model using triplet and quadruplet loss are shown in Table 3. It can be seen that the quadruplet loss in our test settings performed marginally better than the triplet loss, yet it brings an improvement at almost no cost.
4.5 Influence of Input Image Size
Even though the results achieved with 256x128 input images had already sur- passed the current SOTA performance for many backbones (See Table 2), we decided to test if larger images would result in even higher performance. Out- comes of our experiments are presented in Table 4. It can be seen that using larger images allows to further boost performance. In our settings, we achieved best results for input images of size 320x320. Using larger images (480x480) did not bring any advantage. Additional plots are available in the supplement.
4.6 Influence of Re-ranking
One of the methods boosting performance tested in  was re-ranking during test stage. Re-ranking is a post-processing method applied to raw retrieval results in order to improve the accuracy of the results. In our experiments we used k- reciprocal encoding proposed by  similarly to .
Table 5 presents the comparison of RST models with and without using re- ranking. It can be seen that the results with re-ranking post-processing improved significantly for both datasets. The improvement is greatest for mAP and Acc@1, thus suggesting that re-ranking ’pushes’ the positive samples closer to the start of the results list. Such behaviour seems desirable for real-world applications where a visual search user will obtain more relevant items at the top
4.7 Comparison to the State-of-the-art
In Table 6 we compare our results to the state-of-the-art on Street2Shop dataset . We list the best performing ensemble model from the aforementioned paper, as it achieved better results than their best single model. In comparison, we show the results of our best performing single model, which was trained using an RST approach, Resnet50-IBN-A backbone using quadruplet loss during training. It can be seen that generally our approach performs better by a large margin despite using only 320x320 input image size compared to 800x800 used in .
In all categories and in overall performance our model clearly outperforms current state-of-the-art by a large margin. In two categories our model performs marginally worse, but it may be attributed to the fact that we use much smaller images. The categories with inferior performance consist of small items such as belts or eye-wear, where fine-grained details are more important and a high image resolution is beneficial. In our case cropped items are of very small dimensions, which may limit performance.
It is important to note that the mAP, Acc@1, Acc@20 reported by  as overall performance are just average values across all categories for each metric. There, the retrieval stage was limited to products only from a specific category, thus limiting the choice of the model and making retrieval less challenging. We also report the performance in these settings for our model and for  in the Average over categories row of Table 6 to allow fair comparison. Additionally, we propose an unconstrained evaluation, where we conduct retrieval from all gallery images, with no restrictions. Our results are in the Unconstrained retrieval row of Table 6.
The large gap between Average over categories and Unconstrained retrieval derives from the fact that the well performing categories such as skirts and dresses are the most numerous ones, therefore they weigh more in the final results than categories that have few queries, such as belts. Hence, unconstrained retrieval is closer to a weighted average than the simple average named by us as Average over categories, where each category has equal weight when calculating the final results.
In Table 7 results on DeepFashion dataset are presented. Our model outper- forms the results of  in terms of Accuracy for all k values. Especially Acc@1 noted largest relative boost.
4.8 Cross-domain Evaluation
Inspired by the Person Re-Identification domain, where cross-domain evaluation is often performed to measure robustness of the models , , we decided to conduct such a test using our best performing RST model.
Cross-domain evaluation consists of training a model on Dataset A and eval- uation of its performance on Dataset B. In such settings training and test data distributions are different.
In Table 8 we present the results of our cross-domain experiments, both with and without re-ranking. It can be seen that cross-domain performance of the RST model on both datasets is much lower than when model is trained and tested on the same dataset - same domain (See Table 6 and Table 7). We suppose that such a large gap in performance between cross and same-domain evaluation is caused by a significant difference in data distribution of the two datasets. For example DeepFashion contains only three categories: lower, upper and full-body, while Street2Shop contains 11 fine-grained clothing categories.
Even though cross-domain evaluation results are far from our best performing models, they are on a reasonable level when compared to the current state-of-the- art results for each dataset, thus, indicating that the RST model with Resnet50- IBN-A backbone can learn meaningful representation of garments that can be transferred between domains to a large extent.
In this paper we examined the analogies and differences between Person Re- Identification research field and Fashion Retrieval, then examined the transfer- ability of ReID models, and performed adjustments necessary to make them work for fashion items.
Using the approach proposed by  using only global features, a number of improvements and learning optimizations we achieved 54.8 mAP, 72.9 Acc@20 on Street2Shop and 47.3 Acc@1, 79.0 Acc@20 on DeepFashion dataset, establishing new state-of-the-art results for both. The performance on Street2Shop seems particularly robust, compared to previous state-of-the-art as our results were achieved on images several times smaller than the previous top-scoring models.
By analogy to  we consider the results achieved by our model as a strong baseline for further clothes retrieval research. Despite a simple archi- tecture, the baseline model outperforms previous solutions dedicated and cus- tomized specifically for fashion retrieval systems. Our results are a clear indication that field of fashion retrieval can benefit from research established for Per- son Re-Identification. Additionally, we performed supplementary experiments and showed that quadruplet loss may bring further improvements at a negligible cost, and that re-ranking can further boost performance. Finally, we introduced cross-domain evaluation into clothes retrieval research to test robustness of fash- ion retrieval models.
- Chen, W., Chen, X., Zhang, J., Huang, K.: Beyond triplet loss: a deep quadruplet network for person re-identification. arXiv:1704.01719 [cs] (Apr 2017), http:// arxiv.org/abs/1704.01719, arXiv: 1704.01719
- Cho, Y.J., Kim, S.A., Park, J.H., Lee, K., Yoon, K.J.: Joint per- son re-identification and camera network topology inference in multiple cameras. Computer Vision and Image Understanding 180, 34–46 (Mar 2019). https://doi.org/10.1016/j.cviu.2019.01.003, http://www.sciencedirect. com/science/article/pii/S1077314219300037
- Dodds, E., Nguyen, H., Herdade, S., Culpepper, J., Kae, A., Garrigues, P.: Learning Embeddings for Product Visual Search with Triplet Loss and Online Sampling. arXiv:1810.04652 [cs] (Oct 2018), http://arxiv.org/abs/1810.04652, arXiv: 1810.04652
- Ge, Y., Zhang, R., Wu, L., Wang, X., Tang, X., Luo, P.: DeepFashion2: A Versatile Benchmark for Detection, Pose Estimation, Segmentation and Re-Identification of Clothing Images. arXiv:1901.07973 [cs] (Jan 2019), http://arxiv.org/abs/1901. 07973, arXiv: 1901.07973
- Geng, M., Wang, Y., Xiang, T., Tian, Y.: Deep Transfer Learning for Person Re-identification. arXiv:1611.05244 [cs] (Nov 2016), http://arxiv.org/abs/1611. 05244, arXiv: 1611.05244
- Gordo,A.,Almazan,J.,Revaud,J.,Larlus,D.:End-to-endLearningofDeepVisual Representations for Image Retrieval. arXiv:1610.07940 [cs] (May 2017), http:// arxiv.org/abs/1610.07940, arXiv: 1610.07940
- Guo, S., Huang, W., Zhang, X., Srikhanta, P., Cui, Y., Li, Y., Scott, M.R., Adam, H., Belongie, S.: The iMaterialist Fashion Attribute Dataset. arXiv:1906.05750 [cs] (Jun 2019), http://arxiv.org/abs/1906.05750, arXiv: 1906.05750
- Harwood, B., G, V.K.B., Carneiro, G., Reid, I., Drummond, T.: Smart Mining for Deep Metric Learning. arXiv:1704.01285 [cs] (Jul 2017), http://arxiv.org/abs/ 1704.01285, arXiv: 1704.01285
- Hermans, A., Beyer, L., Leibe, B.: In Defense of the Triplet Loss for Person Re- Identification. arXiv:1703.07737 [cs] (Nov 2017), http://arxiv.org/abs/1703. 07737, arXiv: 1703.07737
- Hu, J., Shen, L., Albanie, S., Sun, G., Wu, E.: Squeeze-and-Excitation Networks. arXiv:1709.01507 [cs] (May 2019), http://arxiv.org/abs/1709.01507, arXiv: 1709.01507
- Huang, J., Feris, R., Chen, Q., Yan, S.: Cross-Domain Image Retrieval with a Dual Attribute-Aware Ranking Network. In: 2015 IEEE International Con- ference on Computer Vision (ICCV). pp. 1062–1070. IEEE, Santiago, Chile (Dec 2015). https://doi.org/10.1109/ICCV.2015.127, http://ieeexplore.ieee. org/document/7410484/
- Huang, W., Hu, R., Liang, C., Yu, Y., Wang, Z., Zhong, X., Zhang, C.: Cam- era Network Based Person Re-identification by Leveraging Spatial-Temporal Constraint and Multiple Cameras Relations. In: Tian, Q., Sebe, N., Qi, G.J., Huet, B., Hong, R., Liu, X. (eds.) MultiMedia Modeling. pp. 174–186. Lecture Notes in Computer Science, Springer International Publishing, Cham (2016). https://doi.org/10.1007/978-3-319-27671-7 15
- Jagadeesh, V., Piramuthu, R., Bhardwaj, A., Di, W., Sundaresan, N.: Large Scale Visual Recommendations From Street Fashion Images. arXiv:1401.1778 [cs] (Jan 2014), http://arxiv.org/abs/1401.1778, arXiv: 1401.1778
- Kiapour, M.H., Han, X., Lazebnik, S., Berg, A.C., Berg, T.L.: Where to Buy It: Matching Street Clothing Photos in Online Shops. In: 2015 IEEE Interna- tional Conference on Computer Vision (ICCV). pp. 3343–3351. IEEE, Santiago, Chile (Dec 2015). https://doi.org/10.1109/ICCV.2015.382, http://ieeexplore. ieee.org/document/7410739/
- Kuang, Z., Gao, Y., Li, G., Luo, P., Chen, Y., Lin, L., Zhang, W.: Fashion Retrieval via Graph Reasoning Networks on a Similarity Pyramid. arXiv:1908.11754 [cs] (Aug 2019), http://arxiv.org/abs/1908.11754, arXiv: 1908.11754
- Kucer, M., Murray, N.: A detect-then-retrieve model for multi-domain fashion item retrieval. In: CVPR Workshops (2019)
- Kstinger, M., Hirzer, M., Wohlhart, P., Roth, P.M., Bischof, H.: Large scale metric learning from equivalence constraints. In: 2012 IEEE Confer- ence on Computer Vision and Pattern Recognition. pp. 2288–2295 (Jun 2012). https://doi.org/10.1109/CVPR.2012.6247939, iSSN: 1063-6919
- Liao, L., He, X., Zhao, B., Ngo, C.W., Chua, T.S.: Interpretable Multimodal Re- trieval for Fashion Products. In: 2018 ACM Multimedia Conference on Multime- dia Conference - MM ’18. pp. 1571–1579. ACM Press, Seoul, Republic of Korea (2018). https://doi.org/10.1145/3240508.3240646, http://dl.acm.org/citation. cfm?doid=3240508.3240646
- Liu, J., Zha, Z.J., Tian, Q., Liu, D., Yao, T., Ling, Q., Mei, T.: Multi- Scale Triplet CNN for Person Re-Identification. In: Proceedings of the 24th ACM international conference on Multimedia. pp. 192–196. MM ’16, Asso- ciation for Computing Machinery, Amsterdam, The Netherlands (Oct 2016). https://doi.org/10.1145/2964284.2967209, https://doi.org/10.1145/2964284. 2967209
- Liu, J., Lu, H.: Deep Fashion Analysis with Feature Map Upsampling and Landmark-Driven Attention. In: Leal-Taix, L., Roth, S. (eds.) Computer Vi- sion ECCV 2018 Workshops, vol. 11131, pp. 30–36. Springer International Pub- lishing, Cham (2019). https://doi.org/10.1007/978-3-030-11015-4 4, http://link. springer.com/10.1007/978-3-030-11015-4_4
- Liu, Z., Luo, P., Qiu, S., Wang, X., Tang, X.: DeepFashion: Powering Robust Clothes Recognition and Retrieval with Rich Annotations. In: 2016 IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 1096–1104. IEEE, Las Vegas, NV, USA (Jun 2016). https://doi.org/10.1109/CVPR.2016.124, http://ieeexplore.ieee.org/document/7780493/
- Luo, H., Jiang, W., Gu, Y., Liu, F., Liao, X., Lai, S., Gu, J.: A Strong Baseline and Batch Normalization Neck for Deep Person Re-identification. arXiv:1906.08332 [cs] (Jun 2019), http://arxiv.org/abs/1906.08332, arXiv: 1906.08332
- Quan, R., Dong, X., Wu, Y., Zhu, L., Yang, Y.: Auto-ReID: Searching for a Part- aware ConvNet for Person Re-Identification. arXiv:1903.09776 [cs] (Aug 2019), http://arxiv.org/abs/1903.09776, arXiv: 1903.09776 version: 4
- Razavian, A.S., Azizpour, H., Sullivan, J., Carlsson, S.: CNN Features Off-the- Shelf: An Astounding Baseline for Recognition. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 512–519. IEEE, Colum- bus, OH, USA (Jun 2014). https://doi.org/10.1109/CVPRW.2014.131, http:// ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=6910029
- Sarfraz, M.S., Schumann, A., Eberle, A., Stiefelhagen, R.: A Pose-Sensitive Em- bedding for Person Re-Identification with Expanded Cross Neighborhood Re- Ranking. arXiv:1711.10378 [cs] (Apr 2018), http://arxiv.org/abs/1711.10378, arXiv: 1711.10378
- Schroff, F., Kalenichenko, D., Philbin, J.: FaceNet: A Unified Embedding for Face Recognition and Clustering. 2015 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) pp. 815–823 (Jun 2015). https://doi.org/10.1109/CVPR.2015.7298682, http://arxiv.org/abs/1503. 03832, arXiv: 1503.03832
- Shrivastava, A., Gupta, A., Girshick, R.: Training Region-Based Ob- ject Detectors With Online Hard Example Mining. pp. 761–769 (2016), https://www.cv-foundation.org/openaccess/content_cvpr_2016/html/ Shrivastava_Training_Region-Based_Object_CVPR_2016_paper.html
- Sohn, K.: Improved Deep Metric Learning with Multi-class N-pair Loss Objective. In: NIPS (2016)
- Su, C., Li, J., Zhang, S., Xing, J., Gao, W., Tian, Q.: Pose-driven Deep Convolutional Model for Person Re-identification (Sep 2017). https://doi.org/10.1109/ICCV.2017.427, https://arxiv.org/abs/1709.08325v1
- Tan, M., Le, Q.V.: Efficientnet: Rethinking model scaling for convolutional neural networks. ICML abs/1905.11946 (2019), http://arxiv.org/abs/1905.11946
- Vishvakarma, A.: MILDNet: A Lightweight Single Scaled Deep Ranking Archi- tecture. arXiv:1903.00905 [cs] (Mar 2019), http://arxiv.org/abs/1903.00905, arXiv: 1903.00905
- Wang, G., Lai, J., Huang, P., Xie, X.: Spatial-Temporal Person Re-identification. arXiv:1812.03282 [cs] (Dec 2018), http://arxiv.org/abs/1812.03282, arXiv: 1812.03282 version: 1
- Wang, J., song, Y., Leung, T., Rosenberg, C., Wang, J., Philbin, J., Chen, B., Wu, Y.: Learning Fine-grained Image Similarity with Deep Ranking. arXiv:1404.4661 [cs] (Apr 2014), http://arxiv.org/abs/1404.4661, arXiv: 1404.4661
- Wang, X., Hua, Y., Kodirov, E., Hu, G., Garnier, R., Robertson, N.M.: Ranked List Loss for Deep Metric Learning. arXiv:1903.03238 [cs] (Aug 2019), http:// arxiv.org/abs/1903.03238, arXiv: 1903.03238
- Wen, Y., Zhang, K., Li, Z., Qiao, Y.: A Discriminative Feature Learning Ap- proach for Deep Face Recognition. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision ECCV 2016, vol. 9911, pp. 499–515. Springer Inter- national Publishing, Cham (2016). https://doi.org/10.1007/978-3-319-46478-7 31, http://link.springer.com/10.1007/978-3-319-46478-7_31
- Wu, C.Y., Manmatha, R., Smola, A.J., Krhenbhl, P.: Sampling Matters in Deep Embedding Learning. arXiv:1706.07567 [cs] (Jan 2018), http://arxiv.org/abs/ 1706.07567, arXiv: 1706.07567
- Xingang Pan, Ping Luo, J.S., Tang, X.: Two at once: Enhancing learning and generalization capacities via ibn-net. In: ECCV (2018)
- Yamaguchi, K., Kiapour, M.H., Berg, T.L.: Paper Doll Parsing: Retrieving Similar Styles to Parse Clothing Items. In: 2013 IEEE International Con- ference on Computer Vision. pp. 3519–3526. IEEE, Sydney, Australia (Dec 2013). https://doi.org/10.1109/ICCV.2013.437, http://ieeexplore.ieee.org/ document/6751549/
- Yang, W., Luo, P., Lin, L.: Clothing Co-Parsing by Joint Image Segmentation and Labeling. 2014 IEEE Conference on Computer Vision and Pattern Recognition pp. 3182–3189 (Jun 2014). https://doi.org/10.1109/CVPR.2014.407, http://arxiv. org/abs/1502.00739, arXiv: 1502.00739
- Zhang, K., Zhang, Z., Li, Z., Qiao, Y.: Joint Face Detection and Alignment Us- ing Multitask Cascaded Convolutional Networks. IEEE Signal Processing Letters 23(10), 1499–1503 (Oct 2016). https://doi.org/10.1109/LSP.2016.2603342
- Zhang, X., Luo, H., Fan, X., Xiang, W., Sun, Y., Xiao, Q., Jiang, W., Zhang, C., Sun, J.: AlignedReID: Surpassing Human-Level Performance in Person Re- Identification. arXiv:1711.08184 [cs] (Jan 2018), http://arxiv.org/abs/1711. 08184, arXiv: 1711.08184
- Zheng, L., Bie, Z., Sun, Y., Wang, J., Su, C., Wang, S., Tian, Q.: MARS: A Video Benchmark for Large-Scale Person Re-Identification. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision ECCV 2016. pp. 868–884. Lec- ture Notes in Computer Science, Springer International Publishing, Cham (2016). https://doi.org/10.1007/978-3-319-46466-4 52
- Zheng,S.,Yang,F.,Kiapour,M.H.,Piramuthu,R.:ModaNet:ALarge-ScaleStreet Fashion Dataset with Polygon Annotations. arXiv:1807.01394 [cs] (Jul 2018), http: //arxiv.org/abs/1807.01394, arXiv: 1807.01394
- Zhong, Z., Zheng, L., Cao, D., Li, S.: Re-ranking Person Re-identification with k-Reciprocal Encoding. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3652–3661. IEEE, Honolulu, HI (Jul 2017). https://doi.org/10.1109/CVPR.2017.389, http://ieeexplore.ieee.org/ document/8099872/
- Zhou, K., Yang, Y., Cavallaro, A., Xiang, T.: Learning Generalisable Omni-Scale Representations for Person Re-Identification. arXiv:1910.06827 [cs] (Oct 2019), http://arxiv.org/abs/1910.06827, arXiv: 1910.06827
- Zhou, K., Yang, Y., Cavallaro, A., Xiang, T.: Omni-Scale Feature Learning for Person Re-Identification. arXiv:1905.00953 [cs] (Dec 2019), http://arxiv.org/ abs/1905.00953, arXiv: 1905.00953
- Zou, X., Kong, X., Wong, W., Wang, C., Liu, Y., Cao, Y.: FashionAI: A Hierar- chical Dataset for Fashion Understanding p. 9 (2018)
Mikolaj Wieczorek, Synerise
Andrzej Michalowski, Synerise
Jacek Dabrowski, Synerise [0000−0002−1581−2365],
Anna Wroblewska, Synerise and Faculty of Mathematics and Information Science, Warsaw University of Technology [0000−0002−3407−7570]