The official implementation of the paper:
Under review
Main results on RefCOCO
| Backbone | val | test A | test B |
|---|---|---|---|
| ResNet50 | 73.92 | 76.03 | 70.86 |
| ResNet101 | 74.67 | 76.8 | 70.42 |
Main results on RefCOCO+
| Backbone | val | test A | test B |
|---|---|---|---|
| ResNet50 | 64.02 | 69.74 | 55.04 |
| ResNet101 | 64.80 | 70.33 | 56.33 |
Main results on G-Ref
| Backbone | val | test |
|---|---|---|
| ResNet50 | 65.69 | 65.90 |
| ResNet101 | 66.77 | 66.52 |
update here
We test our work in the following environments, other versions may also be compatible:
- CUDA 11.1
- Python 3.8
- Pytorch 1.9.0
Please refer to installation.md for installation
Please refer to data.md for data preparation.
sh scripts/train.sh
or
python train_net_video.py --config-file <config-path> --num-gpus <?> OUTPUT_DIR <?>
for example, to train Resnet101-backbone model in RefCOCO dataset with 2 gpus:
python train_net_video.py --config-file configs/refcoco/VLFormer_R101_bs8_100k.yaml --num-gpus 2 OUTPUT_DIR output/refcoco-RN101
In terms of resuming the previous training, then add the flag --resume
python train_net_video.py --config-file <config-path> --num-gpus <?> --eval-only OUTPUT_DIR <output_dir> MODEL.WEIGHTS <weight_path>