First, we adopt the Faster-R-CNN-with-model-pretrained-on-Visual-Genome to get the object proposals of unlabeled images. Second, we utilize the bottom-up-attention to get attributes of objects.
-
You can download the detection results from Google Drive and put them in the right place.
$ mv detection_results.tar.gz ./Pseudo-Q/data/ $ tar -zxvf detection_results.tar.gz -
Or you can follow the tutorial in the above two repositories to get detection results by your own. Here, I provide a code script to help researchers to generate object detection and attribute recognition results based on bottom-up-attention.
- Please follow the instruction in bottom-up-attention to build the environment.
- Before running the code, you should get the config file, download the checkpoint weights from Google Drive or Tsinghua Cloud, and put these files on the right path.
- Run the code with following command on the bottom-up-attention codebase.
$ OMP_NUM_THREADS=2 CUDA_VISIBLE_DEVICES=0 python get_detection_results.py --load_dir data/pretrained_model --image_dir /home/data/referit_data/other/images/mscoco/images/train2014 --out_path /cluster/home1/hjjiang/CVPR2022/pseudo_dataset_after_debug_1027/unc/r152_attr_detection_results/ --image_list_file /cluster/home1/hjjiang/CVPR2022/Faster-R-CNN-with-model-pretrained-on-Visual-Genome_2080/statistic/unc/unc_train_imagelist_split0.txt --vg_dataset unc --cuda --split_ind 0
Please download the images of RefCOCO or other dataset, to ../data/image_data first.
$ cd pseudo_sample_generation
$ bash scripts/generate_pseudo_data_unc.sh
The pseudo samples will be stored in ../data/pseudo_sample/.