Skip to content

Latest commit

 

History

History
executable file
·
93 lines (74 loc) · 3.86 KB

README.md

File metadata and controls

executable file
·
93 lines (74 loc) · 3.86 KB
Raw

The SAPNet code is in mmdet/models/detectors/SAPNet.py mmdet/models/roi_heads/SAP_head.py our GUPs: 8 * RTX3090

Prerequisites

#install environment following

conda create -n open-mmlab python=3.9 -y
conda activate open-mmlab
# install latest pytorch prebuilt with the default prebuilt CUDA version (usually the latest)
conda install pytorch==1.9.0 torchvision==0.10.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
or
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.9.0/index.html

pip install mmcv-full ==1.5.0

# install mmdetection

pip uninstall pycocotools   # sometimes need to source deactivate before, for 
pip install -r requirements/build.txt
pip install -v -e . --user  # or try "python setup.py develop" if get still got pycocotools error
chmod +x tools/dist_train.sh
pip install scikit-image

Prepare dataset COCO

  1. download dataset to data/coco
  2. generate point annotation or download point annotation( Baidu Yun passwd:6752 or Google Driver],
  3. generate SAM_results with point prompt move annotations/xxx to data/coco/annotations_qc_pt/xxx

QC Point generation for coco

1.generate QC point annotation

export MU=(0 0)
export S=(0.25 0.25)  # sigma
export SR=0.25 # size_range
export VERSION=1
export CORNER=""
# export T="val"
export T="train"
PYTHONPATH=. python huicv/coarse_utils/noise_data_mask_utils.py "generate_noisept_dataset" \
    "data/coco/annotations/instances_${T}2017.json" \
    "data/coco/coarse_annotations_new/quasi-center-point-${MU[0]}-${MU[1]}-${S[0]}-${S[1]}-${SR}_${VERSION}/${CORNER}/qc_instances_${T}2017_coarse.json" \
    --rand_type 'center_gaussian' --range_gaussian_sigma "(${MU[0]},${MU[1]})" --range_gaussian_sigma "(${S[0]},${S[1]})" \
    --size_range "${SR}"

2.Transfer QC point annotation to 'bbox' and transfer original bbox to 'true_bbox'

the QC point annotation is transfered to 'bbox' with fixed w and h, which is easy for mmdetection reading and dataset pipeline

the original bbox is transfered to 'true_bbox', which is the real box ground-turth

export VERSION=1
export MU=(0 0)
export S=(0.25 0.25)  # sigma
export RS=0.25
export CORNER=""
export WH=(64 64)
export T="train"
PYTHONPATH=. python huicv/coarse_utils/noise_data_utils.py "generate_pseudo_bbox_for_point" \
    "data/coco/coarse_annotations_new/quasi-center-point-${MU[0]}-${MU[1]}-${S[0]}-${S[1]}-${SR}_${VERSION}/${CORNER}/qc_instances_${T}2017_coarse.json"  \
    "data/coco/coarse_annotations_new/quasi-center-point-${MU[0]}-${MU[1]}-${S[0]}-${S[1]}-${SR}_${VERSION}/${CORNER}/qc_instances_${T}2017_coarse_with_gt.json"  \
    --pseudo_w ${WH[0]} --pseudo_h ${WH[1]}

3.For other dataset, we can transform the annotation style to coco json style and use the same way, after this, you can genertae annotation

Train, Test and Visualization

Take COCO as example

Prepare trained model

  1. move coco dataset (2017 version) or make a soft link to data/coco
  2. download weight and annotation(coco and voc2012SBD) from [Baidu Yun(passwd:u5fu)](链接: https://pan.baidu.com/s/1d1sWCpytBlynQ8NUd1O7bw) or Google Driver ,

Train && inference

open to the work path: SAPNet/TOV_mmdetection

  1. SAPNet

    CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=10000 tools/dist_train.sh configs2/COCO/SAPNet/SAPNet_r50_fpn_1x_coco_ms.py 8

  2. inference with trained P2BNet to get pseudo box and train FasterRCNN with pseudo box

    work_dir='../TOV_mmdetection_cache/work_dir/coco/' && CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=10000 tools/dist_test.sh configs2/COCO/P2BNet/SAPNet_r50_fpn_1x_coco.py epoch_12.pth 8