Yisheng (Ethan) He

Yes, the All object adds and add(-s) results are reported. After the MeanShift problem is fixed, we get better add(-s) than the paper-reported results.

It seems like you didn't compile the related code following our README instruction: https://github.com/ethnhe/FFB6D#:~:text=Compile%20RandLA-Net%20operators%3A

That part of code hasn't been reconstructed for open source. I will add it to my to-do list as I'm busy with other projects these days.

No, as illustrated in A.3. Implementation: PVN3D+ICP in our paper, we use the post-processed predicted mask for the whole images, using the fill label function. https://github.com/ethnhe/PVN3D/blob/9a5481a82af4eb89fa8e0ab10ba22b9077b4652e/pvn3d/eval_icp.py#L429 This procedure is slow...

It's a typo and I tried fixing it and re-trained one of the models on LineMOD, finding that the performance is close. Considering the cost of re-training all pre-trained models,...

> I think the upsampling stages should be the 2nd version according to the paper, right? It is strange that you used two final layer before. Yes, but the performance...

- In Table 7 of our paper we can see full flow bidirectional fusion is also beneficial to other frameworks. We also tried CNN with PointNet++ and it also works....

- not simply concatenation, but utilize similar architecture as our point to RGB and RGB to point fusion. - The XYZ map part is the same as described in our...

It requires some modification to deal with multi instances in one scene using our full models: - For data preprocessing, the semantic label is kept still but the center and...

I don't suggest an Open3D version of normal estimation as the normal direction from it points differently on different points, with some "inward" the object surface and some "outward", which...