Awesome

Referring Expression Object Segmentation with Caption-Aware Consistency

PyTorch implementation of our method for segmenting the object in an image specified by a natural language description.

Contact: Yi-Wen Chen (chenyiwena at gmail dot com)

Paper

Referring Expression Object Segmentation with Caption-Aware Consistency <br /> Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin and Ming-Hsuan Yang <br /> British Machine Vision Conference (BMVC), 2019 <br />

Please cite our paper if you find it useful for your research.

@inproceedings{Chen_lang2seg_2019,
  author = {Yi-Wen Chen and Yi-Hsuan Tsai and Tiantian Wang and Yen-Yu Lin and Ming-Hsuan Yang},
  booktitle = {British Machine Vision Conference (BMVC)},
  title = {Referring Expression Object Segmentation with Caption-Aware Consistency},
  year = {2019}
}

Prerequisites

• Python 2.7
• Pytorch 0.2 or 0.3
• CUDA 8.0
• Mask R-CNN: Follow the instructions of the mask-faster-rcnn repo, preparing everything needed for pyutils/mask-faster-rcnn.
• REFER API and data: Use the download links of REFER and go to the foloder running make. Follow data/README.md to prepare images and refcoco/refcoco+/refcocog annotations.
• COCO training set should be downloaded in pyutils/mask-faster-rcnn/data/coco/images/train2014.

Preprocessing

The processed data is uploaded in cache/prepro/.

Training

• <DATASET> <SPLITBY> pairs contain: refcoco unc/refcoco+ unc/refcocog umd/refcocog google

• Output model will be saved at <DATASET>_<SPLITBY>/output_<OUTPUT_POSTFIX>. If there are trained models in this directory, the model of the latest iteratioin will be loaded.

• The iteration when learning rate decay is specified as STEPSIZE in train_*.sh.

1. Train the baseline segmentation model with only 1 dynamic filter:

./experiments/scripts/train_baseline.sh <GPUID> <DATASET> <SPLITBY> <OUTPUT_POSTFIX>

2. Train the model with spatial dynamic filters:

./experiments/scripts/train_spatial.sh <GPUID> <DATASET> <SPLITBY> <OUTPUT_POSTFIX>

3. Train the model with spatial dynamic filters and caption loss:

./experiments/scripts/train_cycle.sh <GPUID> <DATASET> <SPLITBY> <OUTPUT_POSTFIX> att2in2 <CAPTION_LOSS_WEIGHT>

The pretrained Mask R-CNN model should be placed at <DATASET>_<SPLITBY>/output_<OUTPUT_POSTFIX>. If there are multiple models in the directory, the model of the latest iteration will be loaded.

The pretrained caption model should be placed at <DATASET>_<SPLITBY>/caption_log_res5_2/, named as model-best.pth and infos-best.pkl.

4. Train the model with spatial dynamic filters and response loss:

./experiments/scripts/train_response.sh <GPUID> <DATASET> <SPLITBY> <OUTPUT_POSTFIX>

5. Train the model with spatial dynamic filters, response loss and caption loss:

./experiments/scripts/train_cycle_response.sh <GPUID> <DATASET> <SPLITBY> <OUTPUT_POSTFIX> att2in2 <CAPTION_LOSS_WEIGHT>

The pretrained Mask R-CNN model should be placed at <DATASET>_<SPLITBY>/output_<OUTPUT_POSTFIX>. If there are multiple models in the directory, the model of the latest iteration will be loaded.

The pretrained caption model should be placed at <DATASET>_<SPLITBY>/caption_log_response/, named as model-best.pth and infos-best.pkl.

6. Train the model with spatial dynamic filters and response loss for VGG16 and Faster R-CNN:

Download the pre-trained Faster R-CNN model here (coco_900k-1190k.tar), and put the .pth and .pkl files in pyutils/mask-faster-rcnn/output/vgg16/coco_2014_train+coco_2014_valminusminival/

./experiments/scripts/train_vgg.sh <GPUID> <DATASET> <SPLITBY> <OUTPUT_POSTFIX>

Evaluation

1. Evaluate the baseline segmentation model:

./experiments/scripts/eval_baseline.sh <GPUID> <DATASET> <SPLITBY> <OUTPUT_POSTFIX> <MODEL_ITER>

Evaluate the model at <DATASET>_<SPLITBY>/output_<OUTPUT_POSTFIX>, of trained iteration <MODEL_ITER>.

Detection and segmentation results will be saved at experiments/det_results.txt and experiments/mask_results.txt respectively.

2. Evaluate the model with spatial dynamic filters (and caption loss):

./experiments/scripts/eval_spatial.sh <GPUID> <DATASET> <SPLITBY> <OUTPUT_POSTFIX> <MODEL_ITER>

3. Evaluate the model with spatial dynamic filters and response loss (and caption loss):

./experiments/scripts/eval_response.sh <GPUID> <DATASET> <SPLITBY> <OUTPUT_POSTFIX> <MODEL_ITER>

4. Evaluate the model with spatial dynamic filters and response loss for VGG16 and Faster R-CNN:

./experiments/scripts/eval_vgg.sh <GPUID> <DATASET> <SPLITBY> <OUTPUT_POSTFIX> <MODEL_ITER>

Acknowledgement

Thanks for the work of Licheng Yu. Our code is heavily borrowed from the implementation of MattNet.

Note

The model and code are available for non-commercial research purposes only.

• 9/2019: code released