Please know that this branch has not been updated yet. Interested people may use the cleaning branch of the repository.
A PyTorch implementation of the CVPR 2018 paper Domain Adaptive Faster R-CNN for Object Detection in the Wild. The original code used by the authors can be found here.
Ensure all rerequisites mentioned here are satisfied by your machine.
Ensure all images used for training (source) and testing have annotations (in Pascal VOC format).
Modify this line in trainval_net_x.py to the directory of the target dataset.
Changes will be made in the future for ease of use.
Add the source and target datasets (in Pascal VOC format) in the src/ and tar/ folders respectively.
Modify factory.py so that your dataset is usable.
Run training as:
CUDA_VISIBLE_DEVICES=$GPU_ID python trainval_net_x.py \
--src $SOURCE_DATASET_NAME --tar $TARGET_DATASET_NAME \
--da True --adaption_lr True --net res101 \
--bs $BATCH_SIZE --nw $WORKER_NUMBER \
--lr $LEARNING_RATE --lr_decay_step $DECAY_STEP \
--cuda
If you want to evaluate the detection performance of a pre-trained model on your testset run:
python test_net_x.py --dataset $TARGET_DATASET_NAME --net res101 \
--checksession $SESSION --checkepoch $EPOCH --checkpoint $CHECKPOINT \
--cuda
Specify the specific model session, checkepoch and checkpoint, e.g., SESSION=1, EPOCH=5, CHECKPOINT=5931.
Follow instructions given here
This code is built on the python implementation of the Faster-RCNN jwyang/faster-rcnn.pytorch
Part of an ongoing project under Dr. Saket Anand