作者：CPFLAME
项目代码：https://github.com/CPFLAME/centerX
项目介绍：https://zhuanlan.zhihu.com/p/323814368

CenterX

This repo is implemented based on detectron2 and CenterNet

What’s new

• Support imgaug data augmentation
• Support swa
• Support Knowledge Distill, teacher-student, designed by myself
• Support other LR_SCHEDULER
• Support Optimizer RangerLars, not convergence in COCO
• We provide some examples and scripts to convert centerX to Caffe, ONNX and TensorRT format in projects/speedup

What’s comming

• [️✔] Support simple inference
• [✔] Support to caffe, onnx, tensorRT
• Support keypoints

Requirements

• Python >= 3.7
• PyTorch >= 1.5
• torchvision that matches the PyTorch installation.
• OpenCV
• pycocotools

pip install cython; pip install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

• GCC >= 4.9

gcc --version

• detectron2

pip install -U 'git+https://github.com/CPFLAME/detectron2.git'

• pytoch tools

pip install git+https://github.com/pabloppp/pytorch-tools -U

Data prepare

the same as detectron2

Training

modify your yamls in run.sh

sh run.sh

Testing and Evaluation

modify your yamls in run.sh,
add eval-only and MODEL.WEIGHTS in your setting

sh run.sh

Performance

coco

This repo use less training time to get a competitive performance compared to other versions

Backbone ResNet-50

Code	mAP	epoch
centerX	33.2	70
centerX	34.3	140
centernet-better	34.9	140

Backbone ResNet-18

centerX_KD means ResNet-50(33.2) as teacher, ResNet-18(27.9) as student, Knowledge Distill for 70 epoch in coco.

Code	mAP	epoch
centerX	30.2	140
centerX	27.9	70
centerX_KD	31.0	70
centernet-better	29.8	140
centernet	28.1	140

crowd human

• optim: SGD
• lr: 0.02
• scheduler: WarmupMultiStepLR ,drop 0.1 in (50, 62) for 80 epoch; (90 ,120) for 140 epoch
• train size: 512 max size
• test size: 512 max size
• batch size: 64
• woGT: KD only use teacher loss

Backbone	mAP	mAP50	mAP75	epoch	teacher	student_pretrain
resdcn18	31.2	56.6	30.8	80	-	-
resdcn18_swa	31.1	56.6	30.4	80	-	-
resdcn18_syncBN	31.3	56.6	30.7	80	-	-
resdcn18_imgaug	29.6	54.7	28.9	80	-	-
resdcn18_KD	34.5	60.2	34.3	80	resdcn50	resdcn18
resdcn18_KD_woGT	33.0	58.3	32.7	80	resdcn50	resdcn18
resdcn18_KD_woGT_scratch	32.8	58.1	32.6	140	resdcn50	imagenet
resdcn50	35.1	61.2	35.3	80	-	-

KD exp

crowd human KD

Generalization performance for Knowledge Distill

Backbone	crowd mAP	coco_person mAP	epoch	teacher	student_pretrain	train_set
resdcn50	35.1	35.7	80	-	-	crowd
resdcn18(baseline)	31.2	31.2	80	-	-	crowd
resdcn18_KD	34.5	34.9	80	resdcn50	resdcn18	crowd
resdcn18_KD_woGT_scratch	32.8	34.2	140	resdcn50	imagenet	crowd
resdcn18_KD_woGT_scratch	34.1	36.3	140	resdcn50	imagenet	crowd+coco

multi teacher KD

Backbone	mAP crowd	mAP coco_car	epoch	teacher	student_pretrain	train_set
1.resdcn50	35.1	-	80	-	-	crowd
2.resdcn18	31.7	-	70	-	-	crowd
3.resdcn50	-	31.6	70	-	-	coco_car
4.resdcn18	-	27.8	70	-	-	coco_car
resdcn18_KD_woGT_scratch	31.6	29.4	140	1,3	imagenet	crowd+coco_car

Backbone	mAP crowd_human	mAP widerface	epoch	teacher	student_pretrain	train_set
1.resdcn50	35.1	-	80	-	-	crowd
2.resdcn18	31.7	-	70	-	-	crowd
3.resdcn50	-	32.9	70	-	-	widerface
4.resdcn18	-	29.6	70	-	-	widerface
5.resdcn18_ignore_nolabel	29.1	24.2	140	-	-	crowd+wider
6.resdcn18_pseudo_label	28.9	27.7	140	-	-	crowd+wider
7.resdcn18_KD_woGT_scratch	31.3	32.1	140	1,3	imagenet	crowd+wider

License

centerX is released under the Apache 2.0 license.

Acknowledgement

• detectron2
• CenterNet
• CenterNet-better
• CenterNet-better-plus
• FastReID

标签：centerX,KD,--,检测,resdcn18,开源,sh,CenterX,coco
来源： https://blog.csdn.net/weixin_47477271/article/details/110455112