Some problem when show 3d box

[xiaoMrzhang]

I have trained this model and try to show detection result on image. The 2D bounding box seems grate but 3D bounding box looks so bad. I just draw 8 points of 3D bounding box and find some objects used same point. Have you meet same question?

[DMing11]

@xiaoMrzhang I meet the same question, why?

[xiaoMrzhang]

@xiaoMrzhang I meet the same question, why?

I have no idea, may be somethings wrong in output decode.
@maudzung

[maudzung]

Hi @xiaoMrzhang
I think there may no bugs in the decoding process.
As I mention in the pinned issued, 3D bounding boxes are optimized with 1 more step. Please take a look at the issue here. Welcome to contribute to the repo. Thanks!

[xiaoMrzhang]

Thanks for reply @maudzung
I think the 3D keypoint should look like more better even without optomized. And I have visualized the bounding box point and see two car shared one point, that's too strange.

[Owen-Liuyuxuan]

I feel this problem is more serious in images with some cars truncated by the image border.

One part of the code is to match points in vertex from heatmaps to vertex from center+offset.

If some keypoints are truncated by the image border, I feel like the global matching algorithm simply fails.

Maybe we still need to improve the matching steps?

I also have tried training and testing with only offsets and not corners. It can do perfectly in the training set and intuitively does not generalize well to the validation set as good as heatmap-based predictions.

So all in all, imao, the problem could be

• truncation fails the matching algorithms?
• the heatmap literally produces multiple key-points around the same points (as I pointed out in About a fix in "compute_radius" (kitti_data_utils.py) #10)?
• the offset branch does not generalize well and fails the matching algorithm?

[liulingfeng123]

@maudzung
I have some questions I want to ask you, can you add the following contact information? This is my email：[email protected]. looking forward to your reply