Baseline model for Amur Tiger Reid (using cross entropy loss and triplet loss).

The code is modified from ReID-MGN, you can check each folder's purpose by yourself.

Our method is simple and straight-forward: we use ResNet-50 (pretrained on ImageNet) as backbone, and funetune it by using the given training dataset only. To effectively extract part features, we utilize two strategies: 1) uniformly partition in the feature maps; 2) use pose information to divide feature maps.

We modify the cvwc2019 dataset through rotating data that is not conducive to training.
'new_keypoints_train.json' denotes that we adjust keypoints coordinates to fit the input image size. 'new_modified_keypoints_train.json' denotes that we also rotate the coordinates of the corresponding keypoints.

We also use our model to generate pseudo-labels for unlabeled train data, but it`s no time to finetune. 'reid_list_train30_358.csv' means that we use the top 30 percent confidence image ids to group together into one class for every query ids. 'reid_list_train70_427.csv' is in the same way.

You can specify more parameters in opt.py

python main.py --mode train --data_path <path/to/dataset> 

Use pretrained weight or your trained weight

python main.py --mode evaluate --data_path <path/to/dataset> --weight <path/to/weight_name.pt> 

Visualize rank10 query result of one image(query from bounding_box_test)

Extract features will take a few munutes, or you can save features as .mat file for multiple uses

python main.py --mode vis --query_image <path/to/query_image> --weight <path/to/weight_name.pt> 

Ablation Study