Bounding boxes are not coming accurate

[themousepotato]

Hi,

I'm getting the following image as the result on visualizing 3D bounding boxes after evaluation using pretrained models. I've used car_lite. I've evaluated the model on the training set itself. Is there anyone facing similar issues or has anyone come across with a possible workaround? Thanks!

[xmyqsh]

@themousepotato Check your coordinate system transformation, including the lidar, camera and vehicle local system by visualization before training. From the pic you shared, the problem should be on the training side, not just the inference side.

[themousepotato]

@xmyqsh sure, will do. But, I'm not doing any training. I've used the pretrained models given in the README which reproduces the published results. I downloaded from this link from that description.

[xmyqsh]

@themousepotato You can solve this problem with your toe :)

[themousepotato]

@xmyqsh I'm really sorry. I couldn't get much info from your statements. Do you mean that it is a bug in the codebase, I need to fix it and train from scratch? Also, I didn't quite get the testing coordinate system part. Kindly elaborate. That would help me a lot. Thanks.

[themousepotato]

Also, is there any chance that the spconv module is not working properly (at least in my local setup)?

[themousepotato]

ping @xmyqsh

[mtshikomba]

I have a similar issue when using the nuscenes dataset #350

How can I improve this this?

[Kartik17]

Looking from the image, The coordinate system of the training data is not same as the inference data