Skip to content
This repository has been archived by the owner on Mar 4, 2023. It is now read-only.

Latest commit

 

History

History
69 lines (53 loc) · 2.26 KB

TinyPerson.md

File metadata and controls

69 lines (53 loc) · 2.26 KB
Raw

TinyPerson

fixed

  • there are two times which can merge results of sub images: 'during inference' or 'after inference', last version we use 'during inference' policy and keep max_per_img=100, but an full image can have 800+ person. So the right setting is max_per_img=200 for 'after inference' policy, or max_per_img=1000 for 'during inference' policy

1 配置文件

配置dataset和mini_annotations

config相关文件的添加,修改涉及

  • _base_里的dataset
  • num_class/max_per_img && nms_pre
  • anchor scales
  • fix BN requires_grad=False in Backbone (learnable BN/GN is hard for TinyPerson??)
  • adap: neck.start_idx and anchor strides
# dataset
configs2/_base_/datasets/TinyPerson/TinyPerson_detection_640x512.py

# Faster-FPN
configs2\TinyPerson\base\faster_rcnn_r50_fpn_1x_TinyPerson640.py

# RetinaNet
configs2/TinyPerson/base/retinanet_r50_fpn_1x_TinyPerson640.py
configs2/TinyPerson/base/retinanet_r50_fpns4_1x_TinyPerson640.py
configs2/TinyPerson/base/retinanet_r50_fpns4_1x_TinyPerson640_clipg.py

2. performance

All train and test on 2080Ti,

  • CUDA10.1/10.2
  • python3.7, cudatookit=10.2, pytorch=1.5, torchvision=0.6

for Faster-FPN, we think the gain compare to TinyBenchmark may come from the cut and merge during inference running time.

detector num_gpu $AP_{50}^{tiny}$ script
Faster-FPN 4 49.81(1) base/Baseline_TinyPerson.sh:exp1.1
Adap RetainaNet 1 45.85(1) base/Baseline_TinyPerson.sh:exp2.1
Adap RetainaNet 4 46.52(1) base/Baseline_TinyPerson.sh:exp2.2(clip grad)
Adap FCOS 2 47.61(1) base/Baseline_TinyPerson.sh:exp6.1

test time compare

  • base/Baseline_TinyPerson.sh:exp5.1
  • detector: Adap FCOS
run-time crop nms_pre max_per_img max_det $AP_{50}^{tiny}$
Y 1000 100 200 42.93
Y 5000 1000 200 46.11
Y 2000 1000 200 46.11
Y 1000 1000 200 46.11
Y 2000 1000 1000 47.61
N 1000/crop 100/crop 200 45.68
run-time crop nms_pre max_per_img max_det $AP_{50}^{tiny}$
Y 2000 1000 1000 47.61
N 1000/crop 500/crop 1000 46.86
N 1000/crop 100/crop 200 45.68