The implementation of DGCAN is based on MMDetection.
Please refer to get_started.md for installation.
To prepare the dataset,
(1) download Graspnet-1billion.
(2) download our refined rectangle label and views from GoogleDrive.
-- data
-- planer_graspnet
-- scenes
-- depths
-- rect_labels_filt_top10%_depth2_nms_0.02_10
-- views
-- models
-- dex_models
- For training DGCAN, the configuration files are in configs/graspnet/.
python tools/train.py configs/graspnet/faster_r2cnn_r50_1016_rgb_ddd_depth_mh_attention_k.py
CUDA_VISIBLE_DEVICES=0,1 .tools/dist_train.sh configs/graspnet/faster_r2cnn_r50_1016_rgb_ddd_depth_mk_attention_k.py 2
- For testing DGCAN, only support single-gpu inference.
python tools/test_graspnet.py configs/graspnet/faster_r2cnn_r50_1016_rgb_ddd_depth_mk_attention_k.py checkpoints/dgcan_trained.pth --eval graspOur trained checkpoints can be download from GoogleDrive.