the use of aggregated volume about depth-from-motion HOT 4 CLOSED

Good point. We tried before, but it did not further improve the performance.

Because we finally used the concatenation of lifted semantic features (weighted by the predicted depth distribution) and stereo cost volume in the current implementation, I guess the former component plays an important role in the final detection such that the change of stereo cost volume does not affect the performance much.

From another perspective, due to the fluctuation of KITTI performance, we can not get confident conclusions about which implementation is better, which may be better studied on nuScenes and Waymo in the future.

from depth-from-motion.

Thanks for your reply!
To put it another way, the key point is the concatenation of semantic features which are weighted by the distribution of depth.
And the depth distribution is generated from the aggregated volume. To sum up, the essence of monocular compensation is to obtain a better depth distribution, which improves the performance.
I am wondering if I understanded your design correctly.

from depth-from-motion.

Exactly. From our observation of the experimental results, it seems like that. If we only use the stereo volume for the subsequent 3D detection, adding the monocular compensated cost volume may be more important.

from depth-from-motion.

Thanks!

from depth-from-motion.