m-parchami/3dvehicledetection

Unofficial implementation of the "3D Bounding Box Estimation Using Deep Learning and Geometry" paper in Tensorflow.

3dvehicledetection's Issues

Distance error

Hello,

Thank you for your repo, it's well written.

When using your code (and the unofficial implementation), I get quite high distance error compared to the original work.

For example at 20m distance I would get up to 5m error instead of 1m.
From my understanding, the error should be in the compute_center function but I don't quite understand the math behind it.

Do you have idea what could be the issue?

3D bbox estimation gives bad results

Hello,

I trained the model following the notebook but when I test it to generate 3D boxes from 2D labels it gives bad results. It either generates Boxes that are far from the object or with too big dimensions for the target object.
These are examples of results I got knowing that I change average dimensions corresponding to the target object!
One other thing, when running the notebook I found out that coordinates for the 2D bounding box were inversed so I changed (it gave more close result but still not accurate at all)