TransMaS

This repository is the official pytorch implementation of the following paper:

NIPS2021 Mixed Supervised Object Detection by TransferringMask Prior and Semantic Similarity

Yan Liu^∗, Zhijie Zhang^∗, Li Niu^†, Junjie Chen, Liqing Zhang^†

MoE Key Lab of Artificial, IntelligenceDepartment of Computer Science and Engineering, Shanghai Jiao Tong University

Setup

Follow the instructions in Installation to build the projects.

Data

Follow instructions in README.old.md to setup COCO and VOC datasets folder and place the coco and voc files under folder ./datasets. Annotations for the COCO-60, and VOC datasets on Google Drive

• coco60_train2017_21987.json, coco60_val2017_969.json : place under folder ./datasets/coco/annotations/
• voc_2007_trainval.json, voc_2007_test.json: place under ./datasets/voc/VOC2007/

Checkpoints

We provide the model checkpoints of object detection network and MIL classifier. All checkpoint files are on Google Drive, place the files under folder ./output/coco60_to_voc/

Evaluation

The test results of Ours^*(single-scale) on VOC2007 test set in the main paper can be reproduced by executing the following commands:

python -m torch.distributed.launch --nproc_per_node=2 tools/test_net.py --config-file wsod/coco60_to_voc/mil_it0.yaml OUTPUT_DIR "output/coco60_to_voc/mil_it2" MODEL.WEIGHT "output/coco60_to_voc/mil_it2/model_final.pth" WEAK.CFG2 "output/coco60_to_voc/odn_it2/config.yml"

Resources

We have summarized the existing papers and codes on weak-shot learning in the following repository: https://github.com/bcmi/Awesome-Weak-Shot-Learning

Acknowledgements

Thanks to WSOD with Progressive Knowledge Transfer providing the base architecture, iterative training strategy, and data annotations for our project. We further transfer mask prior and semantic similarity to bridge the gap between novel categories and base categories by adding the code for Mask Generator and SimNet.