About filter_bg_thresh

[cainiaoshidai]

Hi, It's excellent work. I found that when running your code, a parameter filter_bg_thresh will cause the output loss to be 0. Effective loss can only be produced if the threshold exceeds this parameter. I don't know if my understanding is correct. Can you tell me what this parameter does？ looking forward to your reply. Best.

[ngoductuanlhp]

Hi @cainiaoshidai, Thanks for your interest in our work! filter_bg_thresh is a soft threshold to filter out points belonging to background classes (floor and wall). The 0 output loss only happens if you train the entire network from scratch. Please refer to our training guideline in which we first pre-train the backbone then train the entire model. Best.

[cainiaoshidai]

Hi @cainiaoshidai, Thanks for your interest in our work! filter_bg_thresh is a soft threshold to filter out points belonging to background classes (floor and wall). The 0 output loss only happens if you train the entire network from scratch. Please refer to our training guideline in which we first pre-train the backbone then train the entire model. Best.

Thanks for your reply. I made a mistake while reading the training guideline. By the way, I want to train in my own dataset. But my dataset doesn't have RGB feature. So whether the RGB feature will have a significant influence on the result of indoor and outdoor datasets. Can you give me some advice? Thanks. Best.

[ngoductuanlhp]

I think the lack of RGB features is not a big problem. However, when adapting to a new dataset, you should fine-tune several hyper-parameters (as suggested in custom_data_guide of SoftGroup) such as the voxel_scale, instance_head_cfg.radius, instance_head_cfg.neighbor in the cfg file. Best.

[cainiaoshidai]

Thanks. I will have a try. Best.