anmyles / pgd

[ECCV 2022] Prediction-Guided Distillation for Dense Object Detection

Prediction-Guided Distillation

PyTorch implementation of our ECCV 2022 paper: Prediction-Guided Distillation for Dense Object Detection

Requirements

• Our codebase is built on top of MMDetection, which can be installed following the offcial instuctions.
• We used pytorch pre-trained ResNets for training.
• Please follow the MMdetection offcial instuction to set up COCO dataset.
• Please download the CrowdHuman and set up the dataset by running this script.

Usage

Set up datasets and pre-trained models

mkdir data
ln -s path_to_coco data/coco
ln -s path_to_crowdhuman data/crowdhuman 
ln -s path_to_pretrainedModel data/pretrain_models 

COCO Experiments

# ------------------------------------
#    Here we use ATSS as an example
# ------------------------------------

# Training and testing teacher model
zsh tools/dist_train.sh work_configs/detectors/atss_r101_3x_ms.py 8
zsh tools/dist_test.sh work_configs/detectors/atss_r101_3x_ms.py work_dirs/atss_r101_3x_ms/latest.pth 8

# Training and testing student model 
zsh tools/dist_train.sh work_configs/detectors/atss_r50_1x.py 8
zsh tools/dist_test.sh work_configs/detectors/atss_r50_1x.py work_dirs/atss_r50_1x/latest.pth 8

# Training and testing PGD model
zsh tools/dist_train.sh work_configs/pgd_atss_r101_r50_1x.py 8
zsh tools/dist_test.sh work_configs/pgd_atss_r101_r50_1x.py work_dirs/pgd_atss_r101_r50_1x/latest.pth 8

CrowdHuman Experiments

# Training teacher, conducting KD, and evalauation
zsh tools/run_crowdhuman.sh

Model Zoo

COCO

Detector	Setting	mAP	Config
FCOS	Teacher (r101, 3x, multi-scale)	43.1	config
-	Student (r50, 1x, single-scale)	38.2	config
-	PGD (r50, 1x, single-scale)	42.5 (+4.3)	config
AutoAssign	Teacher (r101, 3x, multi-scale)	44.8	config
-	Student (r50, 1x, single-scale)	40.6	config
-	PGD (r50, 1x, single-scale)	43.8 (+3.1)	config
ATSS	Teacher (r101, 3x, multi-scale)	45.5	config
-	Student (r50, 1x, single-scale)	39.6	config
-	PGD (r50, 1x, single-scale)	44.2 (+4.6)	config
GFL	Teacher (r101, 3x, multi-scale)	45.8	config
-	Student (r50, 1x, single-scale)	40.2	config
-	PGD (r50, 1x, single-scale)	43.8 (+3.6)	config
DDOD	Teacher (r101, 3x, multi-scale)	46.6	config
-	Student (r50, 1x, single-scale)	42.0	config
-	PGD (r50, 1x, single-scale)	45.4 (+3.4)	config

CrowdHuman

Detector	Setting	MR ↓	AP ↑	JI ↑	Config
DDOD	Teacher (r101, 36 epoch, multi-scale)	41.4	90.2	81.4	config
-	Student (r50, 12 epoch, single-scale)	46.0	88.0	79.0	config
-	PGD (r50, 12 epoch, single-scale)	42.8 (-3.2)	90.0 (+2.0)	80.7 (+1.7)	config

Ciation

@article{yang2022predictionguided,
  title={{Prediction-Guided Distillation for Dense Object Detection}},
  author={Yang, Chenhongyi and Ochal, Mateusz and Storkey, Amos and Crowley, Elliot J},
  journal={ECCV 2022},
  year={2022}
}

Acknowledgement

We thank FGD and DDOD for their code base.