Fashion推荐
2019-Complete the look: Scene-based complementary product recommendation
引用数:20 来源:Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.
bridge the gap between scene images and product images,任务:Learning scene-product compatibility,constructing appropriate groundtruth data。we seek to learn a unified style space where compatible scenes and products are close, as they ought to represent similar styles.
- 以前的研究:product-to-product compatibility、retrieve similar looking (or even identical) products given a scene image。Clothing ‘parsing’、pairwise compatibility、outfit compatibility。
- 数据集:Shop the Look:retrieving visually similar (or even identical) products based on a scene image and a bounding box containing the query product。containing compatible pairs consisting of a scene image and a product image。
- 数据集:human-labeled datasets which include bounding boxes of products in scene images,the associated product images, as well as the category of each product。如Exact Street2Shop、Pinterest’s Shop The Look,
- 本文数据集:to remove the product by cropping the scene image。model ‘sees’ the product in the scene image;learn to detect whether the product appears in the scene,instead of measuring compatibility。
- Generating CTL Datasets:consider four candidate regions (i.e., top, bottom, left, and right) that don’t overlap with bounding box B;select the greatest area as the cropped scene image;discard scene-product pairs for which the area of the cropped image is smaller than a threshold。
- 方法:1)adopt ResNet-50 to extract visual features。a visual feature vector from the final layers (e.g. pool5), and a feature map(w×h) from intermediate convolutional layers (e.g. conv4 6)。2)measure compatibility by considering both global and local compatibility in a unified style space
- Global compatibility:learn style embeddings from compatible scene and product images, where nearby embeddings indicate high compatibility,a two-layer feed forward network transform the visual features to a d-dimensional
- Local compatibility:match every region of the scene image with the product image;adopt category-aware attention to assign weights over all regions。
- 目标函数:hinge loss,(scene , positive product, negative product)negative product from the same category
2019-Dressing as a whole: Outfit compatibility learning based on node-wise graph neural networks
引用数:20 来源:The World Wide Web Conference
represent an outfit as a graph. In particular, each node represents a category and each edge represents interaction between two categories. Accordingly, each outfit can be represented as a subgraph by putting items into their corresponding category nodes.
- 之前工作:pair representation and sequence representation;
- 数据集:Polyvore dataset,21889 outfits covering 380 categories;17316 for training, 1497 for validation and 3076 for testing;修改:categories appear more than 100 times,120 categories remain;outfits consisting of more than 3 items;remain 16983 outfits in the training set, 1497 in the validation set and 2697 in the test set.
- Features:images of items and titles,InceptionV3,filter out the words which appear in less than 5 items and less than 3 characters,2757 words。
- 问题定义:Given an outfit s consisting of |s| items (each item has an image and textual description), we aim to predict the compatibility score of the outfit。
- Graph Construction:if two categories have matching relation in the training dataset, there are two directed edges in reverse directions between the two corresponding nodes.
- Learning Initial Node State:input of NGNN is features (either visual or textual features) of items,map its feature to a latent style space with the size of d.
- Modeling Node Interactions:nodes will receive sum of neighbors’ state。update the state similar to GRU .
- Compatibility Calculating with Attention Mechanism:1)items have different influence on an outfit.2)same item plays different roles in different outfits。utilize self-attention to calculate the graph-level output and also the outfit compatibility score。
- NGNN with Multi-modal Input:input the visual features and textual features to the two channels of NGNN respectively and obtain a visual compatibility score and a textual compatibility score;use a regularization to ensure the consistency between the visual and textual feature of the same item in the shared latent space。
- Training Strategy:outfits in the training dataset as positive outfits,randomly select an item replace it with a random item to form a negative
outfit。 - Task Description:
- Fill-in-the-blank Fashion Recommendation:Given a set of fashion items and a blank, we aim to find the most compatible item from the candidates set to fill in the blank。
- Compatibility Prediction:to predict the compatibility score for any given outfit
2019-Personalized Capsule Wardrobe Creation with Garment and User Modeling
引用数:3 来源:Proceedings of the 27th ACM International Conference on Multimedia.
automatic creation of capsule wardrobes