Baselines detections and tracking for non-keyframes

Baseline detections and tracking results (PointPillars, Megvii and Mapillary) provided within the tracking challenge are really useful for my research and I have two questions in that regard.

1. Would it be possible to get detections and tracking results for the non-keyframes as well?

From the previous discussions on the forum, I have seen that it is possible to use linear interpolation between the keyframes. However, it would be really useful to get uncorrelated detections at the full frame rate.
It would especially be useful for the cameras, since the Mapillary code is not available.
Additionally, from the PointPillars and Megvii papers, it is clear that the lidar detections are based on 10 sweeps, but it would be interesting to compare that approach with a single sweep detection approach.

2. Would it be possible to get access to the tracking results on the train scenes?

In the current files, detections exists for train,val and test scenes. On the other hand, tracking results are given for only val and test scenes.

Hi,

  1. unfortunately not. I don’t have the code for these, but only reached out to the authors to run their method for me on the train/val/test splits. Perhaps you can use the Megvii codebase to reconstruct them ? (https://github.com/poodarchu/Det3D)
    For a comparison of different number of sweeps, take a look at the nuScenes paper, Table 3: https://arxiv.org/pdf/1903.11027.pdf

  2. Unfortunately the authors of https://github.com/xinshuoweng/AB3DMOT only provided us with the tracking results for val and test. I suggest you use their code to extract the results on the training set. In general our baselines weren’t very good (perhaps unsuitable hyperparameter?). I suggest you use the challenge winning tracking method: https://github.com/eddyhkchiu/mahalanobis_3d_multi_object_tracking

1 Like