Skip to content

Latest commit

 

History

History
101 lines (82 loc) · 3.13 KB

README_EN.md

File metadata and controls

101 lines (82 loc) · 3.13 KB
Raw

Detector-in-Detector

Unofficial Implementation of Detector-in-Detector: Multi-Level Analysis for Human-Parts in simplify form.

中文简介

Brief introduction

This project is based on tensorflow object detection framework luminoth, and overload the faster-rcnn interface to implement detector-in-detector in simplify form.

Paper Link: https://arxiv.org/abs/1902.07017

Installtion

train.py
models/fasterrcnn/__init__.py
models/fasterrcnn/detector_in_detector.py
models/fasterrcnn/part_detector.py
models/fasterrcnn/part_detector_rcnn.py
models/fasterrcnn/rcnn_overload.py
models/fasterrcnn/rcnn_proposal_overload.py
models/fasterrcnn/part_detector_roi_pool.py
predict.py
utils/detector_in_detector_predicting.py

Training steps

  • 1 download paper dataset according to
    https://github.com/xiaojie1017/Human-Parts
  • 2 run dataset_script/preprocess_dataset.py to generate luminoth
    friendly csv annotation file
  • 3 generate tfrecord file with luminoth dataset api
  • 4 use luminoth train -c dataset_script/start.yml to train
  • 5 use tensorboard to valid your net.
  • 6 use luminoth predict -c dataset_script/start.yml to predict

Details

How to debug

In train step:
edit utils/hooks/image_vis_hook.py and retrieve in after_run

In predict step:
edit utils/detector_in_detector_predicting.py

Some difference with luminoth

This is only a edit version of luminoth FasterRCNN. the only difference between this yml with ori. faster-rcnn yml is to set main_label_index to differciate body (main_part) bboxes and part bboxes

Some difference with paper

Paper use ground truth person proposals as padding to early training process, here is replaced by random samping and prob threshold filter. You can try to increase the padding num (which is 7) to approximate it.
Or you can switch it by hand.

Some drawbacks may require improved

For the usage of filter script, the data_augmentation config is limited to flip and random noise add action, some patch action may cause main and part overlap be zero, which is harmful for training.
This is related with the construction between Body Detector and Part Detector,If you can accumulate samples and perform train in this fasion, you can tackle this problem.

Lumi faster-rcnn only support batch-size 1, limit the train efficiency.

Not overload lumi eval and web server interface in this version. (may be simple)

Contact

svjack - [email protected] - [email protected]

Project Link:https://github.com/svjack/Detector-in-Detector