Code for Background Learnable Cascade for Zero-Shot Object Detection

Code requirements

• python: python3.7
• nvidia GPU
• pytorch1.1.0
• GCC >=5.4
• NCCL 2
• the other python libs in requirement.txt

Install

conda create -n BLC python=3.7 -y
conda activate BLC

conda install pytorch=1.1.0 torchvision=0.3.0 cudatoolkit=10.0 -c pytorch

pip install cython && pip --no-cache-dir install -r requirements.txt

python setup.py develop

Dataset prepare

• Download the train and test annotations files for BLC from annotations, put all json label file to

  data/coco/annotations/

• Download MSCOCO-2014 dataset and unzip the images it to path：

  data/coco/train2014/
  data/coco/val2014/

• training:

  • train BLRPN:

    • 48/17 split:

      ./tools/dist_train.sh configs/train/BLRPN/blrpn_r50_fpn_semantic_48_17_1x.py 4

    • 65/15 split:

      ./tools/dist_train.sh configs/train/BLRPN/blrpn_r50_fpn_semantic_65_15_1x.py 4

  • generate new Ws:

    • 48/17 split:

      python tools/replace_bg_w2vec_48_17.py

    • 65/15 split:

      python tools/replace_bg_w2vec_65_15.py

  • train Cascade Semantic R-CNN with new Ws：

    • 48/17 split:

        ./tools/dist_train.sh configs/BLC/train/Cascade_Semantic_R-CNN/cascade_semantic_rcnn_information_flow_learnable_bg_48_17_1x.py 4

    • 65/15 split:

      ./tools/dist_train.sh configs/BLC/train/Cascade_Semantic_R-CNN/cascade_semantic_rcnn_information_flow_learnable_bg_48_17_1x.py 4

• Inference & Evaluate:

  • ZSD task:

    • 48/17 split ZSD task:

      • download ms 48/17 BLC ms model, put it in checkpoints/BLC_ms_48_17.pth

      • inference:

        python tools/test.py configs/BLC/inference/zsd/cascade_semantic_rcnn_information_flow_learnable_bg_48_17_1x.py  checkpoints/BLC_ms_48_17.pth --out results/zsd_48_17.pkl

      • our results zsd_48_17.pkl can also downloaded from zsd_48_17.pkl.

      • evaluate:

        python tools/zsd_eval.py results/zsd_48_17.pkl configs/BLC/inference/zsd/cascade_semantic_rcnn_information_flow_learnable_bg_48_17_1x.py

    • 65/15 split ZSD task:

      • download ms 65/15 BLC model, put it in checkpoints/BLC_ms_65_15.pth

      • inference:

        python tools/test.py configs/BLC/inference/zsd/cascade_semantic_rcnn_information_flow_learnable_bg_65_15_1x.py  checkpoints/BLC_ms_65_15.pth --out results/zsd_65_15.pkl

      • our results zsd_65_15.pkl can also downloaded from zsd_65_15.pkl.

      • evaluate:

        python tools/zsd_eval.py results/zsd_65_15.pkl configs/BLC/inference/zsd/cascade_semantic_rcnn_information_flow_learnable_bg_65_15_1x.py

  • GZSD task:

    • 48/17 split GZSD task:

      • download 48/17 BLC model, put it in checkpoints/BLC_48_17.pth

      • inference:

        python tools/test.py configs/BLC/inference/gzsd/gzsd_cascade_semantic_rcnn_information_flow_learnable_bg_48_17_1x.py checkpoints/BLC_48_17.pth --out results/gzsd_48_17.pkl

      • our results gzsd_48_17.pkl can also downloaded from gzsd_48_17.pkl.

      • evaluate:

        python tools/gzsd_eval.py results/gzsd_48_17.pkl configs/BLC/inference/gzsd/gzsd_cascade_semantic_rcnn_information_flow_learnable_bg_48_17_1x.py

    • 65/15 split ZSD task:

      • download 65/15 BLC model, put it in checkpoints/BLC_65_15.pth
      • inference:

        python tools/test.py configs/BLC/inference/gzsd/gzsd_cascade_semantic_rcnn_information_flow_learnable_bg_65_15_1x.py checkpoints/BLC_65_15.pth --out results/gzsd_65_15.pkl

      • our results gzsd_65_15.pkl can also downloaded from gzsd_65_15.pkl.
      • evaluate:

        python tools/gzsd_eval.py results/gzsd_65_15.pkl configs/BLC/inference/gzsd/gzsd_cascade_semantic_rcnn_information_flow_learnable_bg_65_15_1x.py

        License

ZSD is released under MIT License.

Citing

If you use BLC in your research or wish to refer to the baseline results published here, please use the following BibTeX entries:

@InProceedings{Zheng_2020_ACCV,
    author    = {Zheng, Ye and Huang, Ruoran and Han, Chuanqi and Huang, Xi and Cui, Li},
    title     = {Background Learnable Cascade for Zero-Shot Object Detection},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
    month     = {November},
    year      = {2020}
}