Category Dictionary Guided Unsupervised Domain Adaptation for Object Detection

Data Preparation

Public datasets are available from official websites or mirrors. If your folder structure is different, you may need to change the corresponding paths in config files.

scl
├── cfgs
├── lib
├── weights
├── data
│   ├── cityscape
│   │   ├── annos
│   │   ├── annos_multi
│   │   ├── imgs
│   │   ├── train.txt
│   │   ├── val.txt
│   │   ├── train_car.txt
│   │   ├── val_car.txt
│   ├── foggycity
│   │   ├── annos
│   │   ├── annos_multi
│   │   ├── imgs
│   │   ├── train.txt
│   │   ├── val.txt
│   │   ├── train_car.txt
│   │   ├── val_car.txt
|   ├── watercolor
│   │   ├── Annotations
│   │   ├── ImageSets
│   │   │   ├── Main
│   │   │   |    ├── train.txt
│   │   │   |    ├── test.txt
│   │   ├── JPEGImages
|   ├── sim10k
│   │   ├── Annotations
│   │   ├── JPEGImages
│   │   ├── train.txt
│   │   ├── val.txt
│   ├── VOCdevkit
│   │   ├── VOC2007
│   │   │   ├── Annotations
│   │   │   ├── JPEGImages
│   │   │   ├── ImageSets
│   │   ├── VOC2012
│   │   │   ├── Annotations
│   │   │   ├── JPEGImages
│   │   │   ├── ImageSets

Pretrained Model

We used two pretrained models in our experiments, VGG and ResNet101. You can download these two models from:

VGG16: Dropbox, VT Server
ResNet101: Dropbox, VT Server

Download them and put them into the data/pretrained_model/.

CDG model

The trained checkpoints can be accessed by Google drive.

Compilation

As the code is a little old, both pytorch-0.4.0 and pytorch-1.0 (or higher) are needed.

Compile the cuda dependencies using following simple commands:

cd lib
# Under pytorch-0.4.0
sh make_1.sh 
# Under pytorch-1.0
sh make_2.sh

Train

Get the annotator

To train a faster R-CNN model with vgg16 on sim10k, simply run:

 python train_source.py \
            --dataset sim10k --net vgg16 --bs $BATCH_SIZE

For citycape dataset, we first use cyclegan(https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix) to translate the source data into the style of target and then train a source model based on the new transferred source data.

 python train_source.py \
            --dataset cityscape --net vgg16 --bs $BATCH_SIZE

For watercolor dataset, feature alignment by adversarial training is performed to get the annotator.

 python train_adv.py \
            --dataset pascal_voc_0712_water --dataset_t watercolor --net res101 --bs $BATCH_SIZE

Extract training features for learning dictionaries

Extarct foreground features on sim10k:

 python generate_gt_features.py --dataset sim10k --net vgg16 --r True --load_name $PATH_TO_ANNOTATOR

Extarct background features on sim10k:

 python generate_bg_features.py --dataset sim10k  --net vgg16 --r True --load_name $PATH_TO_ANNOTATOR

Learn dictionaries

Learn dictionaries for foreground classes:

 python learn_dicts.py

Learn background dictionaries by changing the settings in the 'learn_dicts.py' file.

Generate pseudo labels for target data

 python generate_inv_matrix.py

 python generate_annos.py --dataset_t cityscape_car --net vgg16 --r True --load_name $PATH_TO_ANNOTATOR

Perform self-training

When doing self-training, the annotations path should be changed in the definition of dataset class accordingly.

 python train_cdg.py --dataset $DATASET_SOURCE --dataset_t $DATASET_TARGET --net $NET --bs $BATCH_SIZE

Test

python test_net.py --dataset $DATASET --net $NET --r True --load_name $PATH_TO_ANNOTATOR

buaali/CDG