Home

Awesome

Localization Distillation for Dense Object Detection

English | 简体中文

Rotated-LD-mmRotate and Rotated-LD-Jittor for rotated object detection are now released.

This repo is based on mmDetection.

Analysis of LD in ZhiHu: 目标检测-定位蒸馏 (LD, CVPR 2022) and 目标检测-定位蒸馏续集——logit蒸馏与feature蒸馏之争

This is the code for our paper:

@Inproceedings{LD,
  title={Localization Distillation for Dense Object Detection},
  author={Zheng, Zhaohui and Ye, Rongguang and Wang, Ping and Ren, Dongwei and Zuo, Wangmeng and Hou, Qibin and Cheng, Ming-Ming},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={9407--9416},
  year={2022}
}

@Article{zheng2023rotatedLD,
  title={Localization Distillation for Object Detection},
  author= {Zheng, Zhaohui and Ye, Rongguang and Hou, Qibin and Ren, Dongwei and Wang, Ping and Zuo, Wangmeng and Cheng, Ming-Ming},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2023},
  volume={45},
  number={8},
  pages={10070-10083},
  doi={10.1109/TPAMI.2023.3248583}}

[2022.12.3] Rotated-LD-Jittor is now available.

[2022.4.13] Rotated-LD-mmRotate is now available.

[2021.3.30] LD is officially included in MMDetection V2, many thanks to @jshilong , @Johnson-Wang and @ZwwWayne for helping migrating the code.

LD is the extension of knowledge distillation on localization task, which utilizes the learned bbox distributions to transfer the localization dark knowledge from teacher to student.

LD stably improves over GFocalV1 about ~2.0 AP without adding any computational cost!

Introduction

Knowledge distillation (KD) has witnessed its powerful capability in learning compact models in object detection. Previous KD methods for object detection mostly focus on imitating deep features within the imitation regions instead of mimicking classification logits due to its inefficiency in distilling localization information. In this paper, by reformulating the knowledge distillation process on localization, we present a novel localization distillation (LD) method which can efficiently transfer the localization knowledge from the teacher to the student. Moreover, we also heuristically introduce the concept of valuable localization region that can aid to selectively distill the semantic and localization knowledge for a certain region. Combining these two new components, for the first time, we show that logit mimicking can outperform feature imitation and localization knowledge distillation is more important and efficient than semantic knowledge for distilling object detectors. Our distillation scheme is simple as well as effective and can be easily applied to different dense object detectors. Experiments show that our LD can boost the AP score of GFocal-ResNet-50 with a single-scale 1x training schedule from 40.1 to 42.1 on the COCO benchmark without any sacrifice on the inference speed.

<img src="LD.png" height="220" align="middle"/>

Installation

Please refer to INSTALL.md for installation and dataset preparation. Pytorch=1.7 and cudatoolkits=11 are recommended.

Get Started

Please see GETTING_STARTED.md for the basic usage of MMDetection.

Train

# assume that you are under the root directory of this project,
# and you have activated your virtual environment if needed.
# and with COCO dataset in 'data/coco/'

./tools/dist_train.sh configs/ld/ld_r50_gflv1_r101_fpn_coco_1x.py 8

Learning rate and batch size setting

lr=(samples_per_gpu * num_gpu) / 16 * 0.01

For 2 GPUs and mini-batch size 6, the relevant portion of the config file would be:

optimizer = dict(type='SGD', lr=0.00375, momentum=0.9, weight_decay=0.0001)
data = dict(
    samples_per_gpu=3,

For 8 GPUs and mini-batch size 16, the relevant portion of the config file would be:

optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
data = dict(
    samples_per_gpu=2,

<font color='red'> Do not set your samples_per_gpu larger than 3! </font>

Feature Imitation Methods

We provide several feature imitation methods, including FitNets fitnet, DeFeat decouple, Fine-Grained finegrain, GI location gibox.

    bbox_head=dict(
        loss_im=dict(type='IMLoss', loss_weight=2.0),
        imitation_method='finegrained'  # gibox, finegrain, decouple, fitnet
    )

Convert model

If you find trained model very large, please refer to publish_model.py

python tools/model_converters/publish_model.py your_model.pth your_new_model.pth

Speed Test (FPS)

CUDA_VISIBLE_DEVICES=0 python3 ./tools/benchmark.py configs/ld/ld_gflv1_r101_r50_fpn_coco_1x.py work_dirs/ld_gflv1_r101_r50_fpn_coco_1x/epoch_24.pth

Evaluation

./tools/dist_test.sh configs/ld/ld_gflv1_r101_r50_fpn_coco_1x.py work_dirs/ld_gflv1_r101_r50_fpn_coco_1x/epoch_24.pth 8 --eval bbox
<details open> <summary>COCO</summary> </details> 
./tools/dist_test.sh configs/ld/ld_gflv1_r101_r18_fpn_voc.py work_dirs/ld_gflv1_r101_r18_fpn_voc/epoch_4.pth 8 --eval mAP
<details open> <summary>PASCAL VOC</summary> </details>

Note:

Pretrained weights

VOC 07+12

GFocal V1

pan.baidu pw: ufc8, teacher R101

pan.baidu pw: 5qra, teacher R101DCN

pan.baidu pw: 1bd3, Main LD R101→R18, box AP = 53.0

pan.baidu pw: thuw, Main LD R101DCN→R34, box AP = 56.5

pan.baidu pw: mp8t, Main LD R101DCN→R101, box AP = 58.4

GoogleDrive Main LD + VLR LD + VLR KD R101→R18, box AP = 54.0

GoogleDrive Main LD + VLR LD + VLR KD + GI imitation R101→R18, box AP = 54.4

COCO

GFocal V1

pan.baidu pw: hj8d, Main LD R101→R18 1x, box AP = 36.5

pan.baidu pw: bvzz, Main LD R101→R50 1x, box AP = 41.1

GoogleDrive Main KD + Main LD + VLR LD R101→R18 1x, box AP = 37.5

GoogleDrive Main KD + Main LD + VLR LD R101→R34 1x, box AP = 41.0

GoogleDrive Main KD + Main LD + VLR LD R101→R50 1x, box AP = 42.1

GoogleDrive Main KD + Main LD + VLR LD + GI imitation R101→R50, box AP = 42.4

GFocal V2

GoogleDrive Main KD + Main LD + VLR LD R101→R50 1x, box AP = 42.7

GoogleDrive | Training log Main KD + Main LD + VLR LD R101-DCN→R101 2x, box AP (test-dev) = 47.1

GoogleDrive | Training log Main KD + Main LD + VLR LD Res2Net101-DCN→X101-32x4d-DCN 2x, box AP (test-dev) = 50.5

For any other teacher model, you can download at GFocalV1, GFocalV2 and mmdetection.

AP Landscape

If you want to draw AP landscape, please replace the relevant files with the files in AP_landscape, and run

# config1 and checkpoint1 correspond to the heads you want to pass through

./tools/dist_test.py config1 config2 checkpoint1 checkpoint2 1

Score voting Cluster-DIoU-NMS

We provide Score voting Cluster-DIoU-NMS which is a speed up version of score voting NMS and combination with DIoU-NMS. For GFocalV1 and GFocalV2, Score voting Cluster-DIoU-NMS will bring 0.1-0.3 AP increase, 0.2-0.5 AP75 increase and <=0.4 AP50 decrease, while it is much faster than score voting NMS in mmdetection. The relevant portion of the config file would be:

# Score voting Cluster-DIoU-NMS
test_cfg = dict(
nms=dict(type='voting_cluster_diounms', iou_threshold=0.6),

# Original NMS
test_cfg = dict(
nms=dict(type='nms', iou_threshold=0.6),