Bridging Cross-task Protocol Inconsistency for Distillation in Dense Object Detection

Introduction

This repository is the official implementation of ICCV2023: Bridging Cross-task Protocol Inconsistency for Distillation in Dense Object Detection. * arxiv * ICCV 2023 open access

News and ToDo List

• [x] [2023-09-01] Release checkpoints and logs
• [x] [2023-08-28] Release paper and code
• [x] [2023-07-14] Accepted by ICCV2023 🎉
• [x] [2023-04-07] Publish initial code

Install

This repo is build on mmdetection 2.28.2

Please refer this link to build the environment (mmcv...).

Then execute the following command to install. git clone https://github.com/TinyTigerPan/BCKD.git cd BCKD pip install -v -e .

Train

For single GPU bash python tools/train.py configs/bckd/bckd_r50_gflv1_r101_fpn_coco_1x.py

For multi GPU bash bash tools/dist_train.sh configs/bckd/bckd_r50_gflv1_r101_fpn_coco_1x.py 8

Eval

For single GPU bash python tools/test.py config_file ckpt_file --eval bbox

For multi GPU bash bash tools/dist_test.sh configs_file ckpt_file 8

Result

| Teacher | Student | Schedule | mAP | download | |-------------|------------|----------|-------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | | GFocal-R50 | 1x | 40.1 | | | GFocal-R101 | GFocal-R50 | 1x | 43.2 | log | ckpt | | | GFocal-R34 | 1x | 38.9 | | | GFocal-R101 | GFocal-R34 | 1x | 42.0 | log | ckpt | | | GFocal-R18 | 1x | 35.8 | | | GFocal-R101 | GFocal-R18 | 1x | 38.6 | log | ckpt |

Cite

@InProceedings{Yang_2023_ICCV, author = {Yang, Longrong and Zhou, Xianpan and Li, Xuewei and Qiao, Liang and Li, Zheyang and Yang, Ziwei and Wang, Gaoang and Li, Xi}, title = {Bridging Cross-task Protocol Inconsistency for Distillation in Dense Object Detection}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)}, month = {October}, year = {2023}, pages = {17175-17184} }