Inferences based on one's own data

[ajxklo]

Hello, your work is excellent. However, I have a small problem. When I use scripts/demo_collision_free_grasps.py to infer my own data, there's a strange phenomenon: it can only determine the grab pose as crossing the table from the side, causing none of them to pass the subsequent collision detection.

[adithyamurali]

Hi @ajxklo one question - are you using instance segmentation to pass in just the object point cloud for inference (not including the background table shown here)?

If you are doing that already, i think the issue is that the threshold is too high and all the best grasps are by the side which are getting filtered out by the collision checker. You can try reducing the threshold. You'll have to make sure the return_topk flag is not passed in, as it'll override the thresholding.

If you are not getting enough grasps, I would also increase the number of grasps sampled.

let me know if that worked

[ajxklo]

你好@ajxklo一个问题——您是否使用实例分割来_仅_传递对象点云进行推理（不包括此处显示的背景表）？

如果你已经这么做了，我认为问题在于阈值太高，导致所有最佳抓取动作都集中在边缘，被碰撞检测器过滤掉了。你可以尝试降低阈值。但要注意，不要return_topk传入标志位，否则它会覆盖阈值设置。__

如果抓取次数不够多，我建议增加抓取次数的采样数量。

如果有效的话请告诉我。

感谢您的回复。是的，我使用了 SAM2 进行实例分割，只传递了目标点云进行推理。我也尝试使用命令python scripts/demo_collision_free_grasps.py --depth_image_path ./GraspGenModels/ourdata/depth.npy --segmentation_mask_path ./GraspGenModels/ourdata/workspace_mask.png --rgb_image_path ./GraspGenModels/ourdata/color.png --camera_intrinsics 606.38 605.96 325.33 251.39 --grasp_threshold 0.4 --gripper_config ./GraspGenModels/checkpoints/graspgen_franka_panda.yml --visualize修改了模型grasp_threshold，但结果仍然和之前一样。

[ajxklo]

你好@ajxklo一个问题——您是否使用实例分割来_仅_传递对象点云进行推理（不包括此处显示的背景表）？

如果你已经这么做了，我认为问题在于阈值太高，导致所有最佳抓取动作都集中在边缘，被碰撞检测器过滤掉了。你可以尝试降低阈值。但要注意，不要return_topk传入标志位，否则它会覆盖阈值设置。__

如果抓取次数不够多，我建议增加抓取次数的采样数量。

如果有效的话请告诉我。

It appears that every predicted gripper pose will cross the tabletop.