This repository contains the code for the paper "Active Object Detection with Knowledge Aggregation and Distillation" accepted at CVPR 2024.
- Python>=3.10.9
- torch>=1.13.1
- torchvision>=0.14.1
- mmcv>= 2.1.0
- mmdet>=3.3.0
- mmengine >=0.10.3
- timm>=0.6.13
- loguru
requirements.txt
file is provided for easy installation of the required packages.
We evaluate our method on the following datasets:
- MECCANO
- 100DOH
- EPIC
- Ego4D
Split: fuqichen1998/SequentialVotingDet
To train the teacher model, run the following command:
# for example [meccano]:
bash tools/dist_train.sh configs/active_object/meccano.py [num_gpus]
To evaluate the student model, run the following command:
# for example [meccano]:
bash tools/dist_test.sh configs/active_object/meccano.py [path_to_checkpoint] [num_gpus]
AP75 | AP50 | AP25 | Models | |
---|---|---|---|---|
Meccano | 14.4 | 28.8 | 36.2 | meccano |
100DOH | 31.2 | 53.9 | 58.9 | 100DOH |
AP | AP50 | AP75 | Models | |
ego4d-swin | 40.5 | 60.6 | 41.9 | ego4d-swin |
ego4d-r50 | 31.4 | 34.6 | 28.9 | ego4d-r50 |
epic-swin | 35.2 | 44.1 | 32.5 | epic-swin |
epic-r50 | 30.2 | 30.1 | 22.5 | epic-r50 |
If you find our work useful in your research, please consider citing our paper:
# TODO
We would like to thank the authors of mmdetection for providing the codebase for object detection.