Issues with KGN Inference Performance and Grasp Point Generation

[Evan-github-account]

I truly appreciate the innovation and effort in your work. I am currently trying to integrate your system into my project, but I have encountered some challenges regarding its inference performance and grasp point generation.

1.Inference Performance
In the demonstration video, it was mentioned that Keypoint-GraspNet achieves 9 FPS. However, during my tests using 1280x720 resolution images, the inference time is approximately 0.6 seconds per frame, which translates to about 1.6 FPS. This is significantly lower than the performance demonstrated in the video.

2.Grasp Point Generation
The grasp points generated during inference seem unusual. Specifically, the grasp locations and scores do not appear to align with the expected results. This issue might be related to model configuration or preprocessing steps, but I have not yet identified the root cause.(Any recommendations to improve the accuracy and reliability of the generated grasp points, including preprocessing steps or key parameters to fine-tune.)

Testing Environment:

GPU: NVIDIA RTX 3090
Image Resolution: 1280x720

Thank you so much for your time and for sharing this valuable work with the community. I look forward to your reply.