Ishan Misra comments

Results 9 comments of


                                            Ishan Misra

bbox_3d_iou, the same two 3d boxs, iou score is zero, zero.why?

We followed prior work and used the VoteNet codebase for our evaluations (https://github.com/facebookresearch/votenet/blob/main/utils/box_util.py#L92) It's quite surprising that it returns 0. Can you check that codebase and see if it returns...

why do you perform down sampling after the first layer in 3detr-m, rather than the whole encoder?

We followed PointNet++ for this design decision, where the downsampling is performed after the first layer. In initial experiments, directly downsampling gave worse results.

Problems with training on my own datasets

`num_semcls` needs to be the number of classes you have (1). It is set to 10 in your code snippet.

Problems with training on my own datasets

`num_angle_bin` is used for datasets like SUN RGB-D when the boxes are rotated. For ScanNet, the boxes are axis-aligned, so the `angle` part is not used. So, if your boxes...

Problems with training on my own datasets

Hi @liudaxia96 I haven't seen this error. It seems to be raise in tensorboardX. Could you try running the code without tensorboard logging?

Questions about AP on scannet

Hi @sunmaosheng755 Apologies for the late response. I have responded to your detailed email.

Different loss function weights between this repo and paper

@ch3cook-fdu thanks for your interest. Which loss function weights are you looking at? The paper screenshot is here: The no_object_weight of 0.25 was more stable across runs, although you can...

Different loss function weights between this repo and paper

Ah, I see what you mean. The weights in the paper are for an older version of the loss that was normalized differently, and so the repo has the right...

Sharing models through the Hugging Face Hub

Hi @osanseviero We'd be happy to share our models there, but it maybe a while since we are busy with a paper deadline. I'll keep this task open for the...