Reid Thesis Exploration
Harmonious Attention Network for Person Re-Identification
focuses on spatial attention and channel attention. The structural design is more ingenious, and the features are refined through two levels of global and local.
reference link
Pose-guided Visible Part Matching for Occluded Person ReID
It consists of three key components: a pose-guided attention (PGA) model for partial feature pooling, a pose-guided visibility predictor (PVP) and a pseudo The features of the label correspond to the model. Three loss functions are used, including Lv, Lm and Lc.
Main method: use the openpose model to perform key point processing in the early stage to generate 18 key point heat maps K and 38 partial affinity fields Lp, and perform occlusion attention discrimination on the basis of the key point heat map, effectively providing visible information for unoccluded parts. Higher score weights reduce the network's feature exploration of occluded parts.
It is necessary to add pose as a guide. Although the author said that end-to-end has been realized, the early processing is essential, and the application layer is still not in place in one step.
PCB：Beyond Part Models: Person Retrieval with Refined Part Pooling
Hard part mode: Extract features by fixedly dividing into 6 part blocks, and finally classify through softmax, in which the features of each part are aligned through the Refined part pooling (RPP) attention mechanism, and finally merge 6 when doing similarity processing A part feature is calculated.
It is not friendly to occluded human figures, and the hard division method is too violent for the overall characteristics. Even if the internal part alignment is done through the soft attention mechanism, it is difficult to distinguish between normal and occluded objects.
Learning Discriminative Features with Multiple Granularities for Person Re-Identification
CloudWalk's MGN is similar to part feature extraction, but it is compatible with global and local features. It only uses softmaxLoss for local features, and uses softmaxLoss + triple-loss for global features. Reference:
https://blog.csdn.net/qidailiming1994/article/ details/104578427
Bag of Tricks and A Strong Baseline for Deep Person Re-identification (Megvii)
is a masterpiece of Megvii, the baseline of reid, which provides a lot of effective training tricks, which is very effective for increasing points. The backbone uses ResNet50, and the overall model is slightly big.
Favorite
http://www.404886.com/cms/72550