Localization Distillation for Object Detection

LD

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 ~0.8 AP and ~1 AR100 without adding any computational cost!

Knowledge distillation (KD) has witnessed its powerful ability in learning compact models in deep learning field, but it is still limited in distilling localization information for object detection. Existing KD methods for object detection mainly focus on mimicking deep features between teacher model and student model, which not only is restricted by specific model architectures, but also cannot distill localization ambiguity. In this paper, we first propose localization distillation (LD) for object detection. In particular, our LD can be formulated as standard KD by adopting the general localization representation of bounding box. Our LD is very flexible, and is applicable to distill localization ambiguity for arbitrary architecture of teacher model and student model. Moreover, it is interesting to find that Self-LD, i.e., distilling teacher model itself, can further boost state-of-the-art performance. Second, we suggest a teacher assistant (TA) strategy to fill the possible gap between teacher model and student model, by which the distillation effectiveness can be guaranteed even the selected teacher model is not optimal. On benchmark datasets PASCAL VOC and MS COCO, our LD can consistently improve the performance for student detectors, and also boosts state-of-the-art detectors notably.

Installation

Please refer to INSTALL.md for installation and dataset preparation.

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_gflv1_r101_r50_fpn_coco_1x.py 8

Learning rate 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,

Convert model

After training with LD, the weight file .pth will be large. You'd better convert the model to save a new small one. See convert_model.py#L38-L40, you can set them to your .pth file and config file. Then, run

python convert_model.py

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

COCO 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

GFocalV1 with LD

GFocalV1 with Self-LD

GFocalV2 with LD

VOC Evaluation

./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

GFocalV1 with LD

This is an example of evaluation results (R-101→R-18).

+-------------+------+-------+--------+-------+
| class       | gts  | dets  | recall | ap    |
+-------------+------+-------+--------+-------+
| aeroplane   | 285  | 4154  | 0.081  | 0.030 |
| bicycle     | 337  | 7124  | 0.125  | 0.108 |
| bird        | 459  | 5326  | 0.096  | 0.018 |
| boat        | 263  | 8307  | 0.065  | 0.034 |
| bottle      | 469  | 10203 | 0.051  | 0.045 |
| bus         | 213  | 4098  | 0.315  | 0.247 |
| car         | 1201 | 16563 | 0.193  | 0.131 |
| cat         | 358  | 4878  | 0.254  | 0.128 |
| chair       | 756  | 32655 | 0.053  | 0.027 |
| cow         | 244  | 4576  | 0.131  | 0.109 |
| diningtable | 206  | 13542 | 0.150  | 0.117 |
| dog         | 489  | 6446  | 0.196  | 0.076 |
| horse       | 348  | 5855  | 0.144  | 0.036 |
| motorbike   | 325  | 6733  | 0.052  | 0.017 |
| person      | 4528 | 51959 | 0.099  | 0.037 |
| pottedplant | 480  | 12979 | 0.031  | 0.009 |
| sheep       | 242  | 4706  | 0.132  | 0.060 |
| sofa        | 239  | 9640  | 0.192  | 0.060 |
| train       | 282  | 4986  | 0.142  | 0.042 |
| tvmonitor   | 308  | 7922  | 0.078  | 0.045 |
+-------------+------+-------+--------+-------+
| mAP         |      |       |        | 0.069 |
+-------------+------+-------+--------+-------+
AP:  0.530091167986393
['AP50: 0.759393', 'AP55: 0.744544', 'AP60: 0.724239', 'AP65: 0.693551', 'AP70: 0.639848', 'AP75: 0.576284', 'AP80: 0.489098', 'AP85: 0.378586', 'AP90: 0.226534', 'AP95: 0.068834']
{'mAP': 0.7593928575515747}

Note:

• For more experimental details, please refer to GFocalV1, GFocalV2 and mmdetection.
• According to ATSS, there is no gap between box-based regression and point-based regression. Personal conjectures: 1) If xywh form is able to work when using general distribution (apply uniform subinterval division for xywh), our LD can also work in xywh form. 2) If xywh form with general distribution cannot obtain better result, then the best modification is to firstly switch xywh form to tblr form and then apply general distribution and LD. Consequently, whether xywh form + general distribution works or not, our LD benefits for all the regression-based detector.

Pretrained weights

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

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, <=0.4 AP50 decrease and <=1.5 FPS 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),

Citation

If you find LD useful in your research, please consider citing:

@Article{zheng2021LD,
  title={Localization Distillation for Object Detection},
  author= {Zhaohui Zheng, Rongguang Ye, Ping Wang, Jun Wang, Dongwei Ren, Wangmeng Zuo},
  journal={arXiv:2102.12252},
  year={2021}
}

GitHub