Domain Adaptive Faster R-CNN

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.

This implementation is built on jwyang/faster-rcnn.pytorch. You may find it to be helpful for debugging.

Usage

Ensure all prerequisites mentioned here are satisfied by your machine. This model is meant for pytorch 0.4.0.

Install python dependencies with pip install -r requirements.txt.

Ensure all images used for training (source) and testing have annotations (in Pascal VOC format).

Refer to dummy formatting in ./lib/datasets/pascal_voc.py files to make your dataset usable.

Changes will be made in the future for ease of use.

Compile

Compile cuda dependencies with cd lib && sh make.sh. Incase cuda is not on $PATH use make2.sh.

Train

Add the source and target datasets (in Pascal VOC format) in the src/ and tar/ folders respectively.

Modify factory.py as given in the dummy formatting so it is usable.

Run training as:

CUDA_VISIBLE_DEVICES=$GPU_ID python new_trainval.py \
                   --dataset $DATA SET NAME\
                   --adaption_lr True --net res101 \
                   --bs $BATCH_SIZE --nw $WORKER_NUMBER \
                   --lr $LEARNING_RATE --lr_decay_step $DECAY_STEP \
                   --cuda

Test

For testing your model, ensure the model is placed in the target dataset folder. Run:

python test_net.py --dataset $TARGET_DATASET_NAME --net res101 \
                   --checksession $SESSION --checkepoch $EPOCH --checkpoint $CHECKPOINT \
                   --cuda --load_dir $path/to/dir

Specify the specific model session, checkepoch and checkpoint, e.g., SESSION=1, EPOCH=5, CHECKPOINT=5931.

Demo

Follow instructions given here

Benchmarks

---

Source dataset	Target dataset
Cityscapes	Foggy Cityscapes

Model	Mean AP	Person AP	Rider AP	Car AP	Truck AP	Bus AP	Train AP	Motorcycle AP	Bicycle AP
Faster R-CNN	18.8	17.8	23.6	27.1	11.9	23.8	9.1	14.4	22.8
Y.Chen et al	27.6	25.0	31.0	40.5	22.1	35.3	20.2	20.0	27.1
dafrcnn-pytorch	28.6	25.7	40.1	36.0	20.1	33.9	17.3	24.3	31.3

---

Source dataset	Target dataset
Zoo_Leipzig_Individuen (Chimpanzees)	(Name to be updated) (Monkeys)

Model	Mean AP (Face)
dafrcnn-pytorch	89.1
Faster RCNN	88.2

---

Source dataset	Target dataset
KITTI	Cityscapes

Model	Mean AP (Car)
dafrcnn-pytorch	39.7
Faster RCNN (VGG16)	35.3

---

This particular experiment was with temporal data and is not publically available. Uses camera trap images. Consists of images from Chitals, Tigers, Sambar and Leopards.

Source dataset	Target dataset
Species_2014	Species_2018

Model	Mean AP	tiger AP	sambar AP	leopard AP	chital AP
dafrcnn-pytorch	83.7	79	83.1	90.7	82.2
Faster RCNN	81.6	77.9	83.3	80.4	81.6

---

Comparing results for this paper

Source dataset	Target dataset
Pascal VOC 2007 + 2012	Clipart1K

Model	Mean AP
dafrcnn-pytorch	30.6
Faster RCNN	28.3

---

Source dataset	Target dataset
SIM10K	Cityscapes
dafrcnn-pytorch	34.3
Faster RCNN	30.6

Current Work

Working on a under-segmentation problem observed in our target data set. Example image:

Acknowledgements

This code is built on the python implementation of the Faster-RCNN jwyang/faster-rcnn.pytorch

Completed as an independent study under Dr. Saket Anand