DEEPLY SUPERVISED SELF-ATTENTION LEARNING MODEL FOR PERSON RE-IDENTIFICATION
We proposed a modified model based on learning self-attetnion module divided into spatial and channel attention which shows a better improvement in the generic classfication tasks. Our Current experiments is carried out on the Person-ReID tasks which is framed as a classification problem. The code here is mainly based on the work in Parameter-Free Spatial Attention Network for Person Re-Identification and PCB. This code is implementain of the paper entitled DEEPLY SUPERVISED SELF-ATTENTION LEARNING MODEL FOR PERSON RE-IDENTIFICATION.
As can be seen from the above model , the re-id task here is framed as classifiaction. Feature is extracted using the main backbone model( shown with light blue color above). Our deep self-attention is embedded before the intermediate supervision shown again with light blue and marked as DSA-1, DSA-2 and DSA-3 . The intermediate attention learning is shown in browan color. As in the earlier work by The brown boxes shows 6 part classifiers (P). It only appears in the ablation study. Then the total loss is the summation over all deep supervision losses, six part losses and the loss from the backbone. Notice that our DSP is applied to each of the first three blocks of the main model.
In the below diagram , DSP attention module is presented. Our attention module which is broken into spatial and channel is parametrized which induces cooperative and competitive character among feature during training. The DSP module is implemented in the code inside the file reSnet.py and here is a short code fragment
gate_chan = 1. + torch.tanh(embedding * norm + self.beta) gate_chan is a channel wise feature aggreagation and normalization , embedding is aggregated feature , norm is normalized feature
gate_spa = 1. + torch.tanh(embedding_spa*norm_spa + self.beta_sp) gate_spa is a spatial wise feature aggreagation and normalization
see more detial in the code.
Python 3.6, Pytorch 2.0
use the below command to run by passing command line argument to the main.py , inlcuding dataset path , and other parameters
!python main.py -d market -b 48 -j 4 --epochs 50 --log logs/market/ --combine-trainval --step-size 40 --data-dir Market-1501
Running the code might display extra information other than epoch no, and lose. I put those print function to monitor some code but you may comment them all. I run it on colab directly but you can try it on you local machine if you have GPU.
To extract feature map. Use single_images_features_3.py and single_features_cam.py . Make sure to specify path to the image and the trained model. Note that image from which you want to extract feature need to be converted to .npy array before passed on to the extractor. I already have converted some of the sample image and you see some result in the folder named 'sample_features'
Market-1501 Dataset is avaialable here and DukeMTMC-reID is availabel here
Hereunder we show feature extracted by the original Resnet-50 model and the our DSP model . For each pair of images. The one on the left is the original and the one on the right is our model's feature. (Refer that paper for detail discussion
----------------- Block -1 ------------|---------------------- Block -2 -----------
To left side is correlation of feature learned by our model and the one on the left is from the baseline model. As can be seen, feature learned by our model tend to be much correlated as evidenced from the dense region around the diagonal
If you find this code or part of this helpful for you reserach , please cite our paper "DEEPLY SUPERVISED SELF-ATTENTION LEARNING MODEL FOR PERSON RE-IDENTIFICATION", and paper Xiong Dei and his coauthors for his kind contribution , More in : https://github.com/XiongDei/Spatial-Attention