Dmytro Mykheievskyi, Dmytro Borysenko and Viktor Porokhonskyy
Samsung Research&Development Institute Ukraine (SRK)
SRK_ODESA online tracking solution follows tracking-by-detection approach. The main differentiation point of this solution is the utilization of in-house designed embedding function named ODESA, which offers good generalization and improved similarity estimation.
| Submit # | MOTA, %. | MOTP, %. | FAF, %.. | MT, %. | ML, %. |....FP......|.....FN......| .ID Sw. |
| .......1....... |.....54.9......|.....79.1......|....5.5......|..32.7....|...20.4...| 24609. | 225292 | ..2614. |
| .......2........|.....54.8......|.....79.0......|....7.5......|..35.4....|...19.2...| 33814. | 215572 | ..3750. |
(disclaimer: correct table formatting is only guarantied for Mozilla Firefox 68.0 under 64-bit Windows 7)
The predictor was derived from SRK_ODESA detection solution, which is based on Faster R-CNN with ResNet-101 backbone extended with FPN and 3-stage refining Cascade. By removing RPN block from the detector and restricting its input to the following items: images, boundingboxes from the public set of detects, and the boundingboxes derived (by the predictor) from the public set of detects the predictor design similar to Tracktor ("Tracking without bells and whistles", arXiv: 1903.05625) was obtained. Our predictor fully complies with the first paragraph of Section 2 and the last subsection ('Public detections') of Section 3 of the above publication. However, the public detects were not passing the regressor head at the frame, which they were assigned to. No association between boundingboxes was taking place at this stage.
ODESA embedding employed in this solution was obtained by means of deep metric learning. Within this solution the model was set to produce the embedding of size set to 128 from image patches converted into HSV color space. The same model was applied within SRK_ODESA solution submitted to KITTI Pedestrian Tracking Benchmark. Neither MOT nor KITTI data were used for model training. The details about embedding preparation sufficient to reproduce its properties are going to be provided in the upcoming publication.
In an attempt to reduce FN count our second submit to the challenge was produced by a modified solution. It was extended with ODESA-based classifier and introduced modifications to the predictor settings. At the moment, it is not clear whether the details on the classifier preparation are going to be shared or not. From this point of view the results of our initial submit are easier to reproduce.
Incoming detects were matched using Hungarian algorithm applied in two stages. At the first stage L2-distance between ODESA descriptors was employed as the matching criterion for candidates, which passed Kalman filter gating. It accounted for 99.65% of association events. At the final association stage IoU criterion was applied for remaining unmatched detects. No gating was applied here.
CPU: 3GHz, 1 core; GPU: 1.5Ghz
June 10, 2019 (3 months ago)
June 16, 2019 at 00:00:00 CET
Project page / code:
D. Borysenko, D. Mykheievskyi, V. Porokhonskyy. ODESA: Object Descriptor that is Smooth Appearance-wise for object tracking tasks. In (to be submitted to ECCV'20), .
|54.8||79.0||7.5||35.4 %||19.2 %||33,814||215,572||3,750||5,493||CPU: 3GHz, 1 core; GPU: 1.5Ghz||Public|
|IDF1||ID Precision||ID Recall|