Experiments on GraspAnything++

[vineet2104]

Dear Authors, I had a couple of questions regarding the code.

1. The current dataset loader - GraspAnythingDataset in data/grasp_anything_data.py only covers RGB, Masks and GT Grasp rectangles. Your paper mentions that you utilize grounding dino for generating bounding boxes using language followed by passing these boxes to GraspSAM for grasp detection. Can you share the code files for that as well? I would like to reproduce your results on the GraspAnything++ dataset.

2. I tried to train GraspSAM on the GraspAnything dataset. As per my understanding from model/utils.py, you use the ground truth object masks to sample points for input to GraspSAM. Is that correct?

Thank you.

[dinnno]

1. We utilized the GraspAnything++ code and data available in the Language-driven Grasp Detection repository (https://github.com/Fsoft-AIC/LGD).

2. Your analysis is correct. We used the provided masks from the benchmark to generate points and boxes, which were then used as input for GraspSAM.

Let us know if you have any further questions!
Thank you.

[vineet2104]

The dataset class you have shared in ./data/grasp_anything_data.py does not include functions to load the text instructions from the dataset. Also, I could not find any references to Grounding DINO in your codebase which, as per my understanding, you have used during training for language driven grasp detection. Could you please share those code files as well?

The authors of the Language-driven Grasp Detection repository recently admitted that their training was done on a much larger dataset (Fsoft-AIC/LGD#15 (comment)), for which they have not released any details. And there is a lot of confusion about their data splits. I'd really appreciate if you can confirm if you were able to replicate their results on the GraspAnything++ dataset during your experiments.

[Ekeeer]

@vineet2104 Did you take a try? We failed to reproduce the results of GraspSAM on GraspAnything++. Could you share something about your experiments? Thx.