Skip to content

Latest commit

 

History

History
134 lines (119 loc) · 6.75 KB

README.md

File metadata and controls

134 lines (119 loc) · 6.75 KB
Raw

A Detectron2 Implementation of SAPNetV2

Code release for the paper:
The study of unsupervised domain adaptation for object detection using spatial attention pyramid networks

Overall Architecture

Overall framework of proposed method. The detector should be pretrained on the source domain, both domain data and detector including domain classifier are involved in the domain adaptation stage. In test phase, detector excluding domain classifier infers on target domain test set. Source domain data in the domain adaptation stage can be replaced with target-like source domain images which are source domain images with target domain style generated by Cycle GAN.

Training Framework

Detailed training framework of the proposed method, RPN and RoI loss are supervised loss of normal faster rcnn. For each iteration, the source and target domain batches will be successively sent into the backbone network, and the next batch will be input only if the previous batch has finished computing. When the features are sent to the domain classifier, the domain labels will be given according to which domain the features come from.

Environment

python 3.8.11
pytorch 1.9.0 (conda)
torchvision 0.10.0 (conda)
detectron2 0.5+cu111 (pip)
tensorboard 2.6.0 (pip)
opencv-python (pip)
pycocotools (pip)

Directory Structre

The directory structure should be as follows:

SAPNetV2/
├── configs/        # configuration files
├── datasets/       # datasets
├── detections/     # main implementation
|   ├── da_heads/   # domain classfier, MS-CAM
|   ├── data/       # data registeration
|   ├── evaluation/ # poscal voc metric
|   ├── layers/     # intermediate neural layer
|   ├── meta_arch/  # faster rcnn
|   ├── modeling/   # rpn, MEDM Loss
|   ├── hook.py     # evaluation extension
|   └── trainer.py  # supervised and unsupervised domain adaptation trainer, grad-cam
|
├── tools/          # train object detector
|   └── train_net.py
├── pretrained      # pretrained models
├── outputs/        # training logs
├── test_images/    # inference outputs
└── scripts/        # useful scripts

Data Preparation

  1. make your dataset format to voc or coco, or other format that detectron2 supports
  2. register your dataset at detection/data/register.py
  3. test set format must be VOC becuase we use VOC metric to evaluate result
# VOC format
dataset_dir = $YOUR_DATASET_ROOT
classes = ('person', 'two-wheels', 'four-wheels') # dataset classes
years = 2007
# to find image list at $YOUR_DATASET_ROOT/ImaegSets/Main/{$split}.txt, only "train", "test", "val", "trainval"
split = 'train'
# call your dataset by this
meta_name = 'itri-taiwan-416_{}'.format(split)
# call register_pascal_voc to register
register_pascal_voc(meta_name, dataset_dir, split, years, classes)

Hyper Parameters

Full configuration explanation is here, hyper parameter we can tuned:

  1. MODEL.DA_HEAD.WINDOW_SIZES and MODEL.DA_HEAD.WINDOW_STRIDES, both are parameter of spatial pooling in domain classfier
  2. MODEL.DA_HEAD.LOSS_WEIGHT, domain classifier loss weight, [0, 1.0]
  3. MODEL.DA_HEAD.TARGET_ENT_LOSS_WEIGHT, entropy loss weight, [0, 1.0]
  4. MODEL.DA_HEAD.TARGET_DIV_LOSS_WEIGHT, diversity loss weight, [-1.0, 0]

The Best Perfomance of Hyper Paramenters

source domain images are target-like source domain images which are generated by Cycle GAN.

Scenarios Domain Classifier Loss Weight Entropy Loss Weight Diversity Loss Weight mAP
cityscapes -> foggy cityscapes 0.6 1.0 -0.2 49.63
sim10k -> cityscapes 0.1 0.8 -0.3 49.48
kitti -> cityscapes 0.1 1.0 -0.05 45.79
  1. you can check these training logs in the pretrained/*-best/ using tensorboard --logdir $LOG_PATH
  2. pretrained/*-baseline store the model pretrained with source domain

Usages

  1. Train a normal faster rcnn using source domain data (source domain only)
  2. Train whole model including domain calssfier using baseline model weight
  3. scripts stores sample scripts for three domain adaptation scenarios
basic usage
  • train a model
python tools/train_net.py --config-file $CONFIG_FILE_PATH --num-gpus 1
  • test the model
# update MODEL.WEIGHTS by command line
python tools/train_net.py --config-file $CONFIG_FILE_PATH --num-gpus 1 --eval-only MODEL.WEIGHTS $MODEL_WEIGHT_PATH
  • predict boxes on the test set
python tools/train_net.py --config-file $CONFIG_FILE_PATH --num-gpus 1 --test-images MODEL.WEIGHTS $MODEL_WEIGHT_PATH  MODEL.ROI_HEADS.SCORE_THRESH_TEST 0.75

Script Usage

Please use these scripts at SAPNetV2/

# or sh scripts/../....sh
..../SAPNetV2$ source scripts/../....sh
  • scripts/*/train_source_only.sh is to train object detector using source domain data only, output directory is under outputs/
  • scripts/*/train_sapv2.sh is to train object detector with domain adaptation, output directory is under outputs/
  • scripts/*/test_img.sh is to predict boxes on the test set, output directory is under test_images/
  • scripts/*/eval.sh is to evaluate object detector with and without domain adaptation on the test set
  • scripts/*/attention_mask.sh is to visualize spatial attention mask of domain classfier, output directory is under test_images/
  • scripts/*/grad_cam_domain_calssifier.sh is to visualize class activation map of domain classifier using grad cam, output directory is under test_images/
  • scripts/*/grad_cam_object_detection.sh is to visualize class activation map of object detector using grad cam, output directory is under test_images/

Ablation

Configuration files for ablation are under configs/*/ablation/

  • scripts/*/cyclegan.sh is to train baseline method using souce domain images with target domain style generated by cycle gan
  • scripts/*/medm.sh is to train baseline method with MEDM loss which is implemented in here
  • scripts/*/mscam.sh is to train baseline method with MS-CAM which is implemented in here

Acknowledgement