Research archive / RFS_Filters
From probabilistic filters to real driving scenes: keeping object identities consistent through missed detections, crossings, and changing class labels.
Recorded in the 0.761 evaluation log, October 12, 2022. These are archived validation results; the experiment has not been rerun for this page.
01 / The approach
A detector produces boxes for each frame. A tracker must decide which boxes belong to existing objects, which indicate new arrivals, and which observations to reject. RFS models also represent uncertainty in how many objects exist.
Represents objects that may exist but have not yet been detected. The birth model introduces the possibility of new objects.
Each candidate track carries a probability of existence and a distribution over its state. A missing detection reduces confidence without necessarily ending the track.
A PMBM filter retains a mixture of association hypotheses. The repository’s PMBMGNN variant propagates the highest-weight global hypothesis to reduce that branching.
Boxes, class labels, scores, and velocity estimates in the nuScenes format.
Propagate position, velocity, survival, and possible object births.
Gate candidate matches, update tracks, and manage alternative assignments.
Export identities and boxes, calculate metrics, and examine scene-level errors.
Implementation references: nuScenes runner and PMBMGNN procedure.
02 / Recorded results
The attached log explicitly identifies the validation split and 6,019 samples. It does not establish the exact detector checkpoint or tracker configuration, so this result is presented as an archived run rather than a controlled comparison.
| Class | AMOTA ↑ | AMOTP ↓ (m) | ID switches |
|---|---|---|---|
| Bicycle | 0.674 | 0.663 | 2 |
| Bus | 0.873 | 0.523 | 9 |
| Car | 0.861 | 0.339 | 420 |
| Motorcycle | 0.823 | 0.532 | 10 |
| Pedestrian | 0.831 | 0.414 | 131 |
| Trailer | 0.529 | 0.936 | 2 |
| Truck | 0.739 | 0.528 | 28 |
The log reports 602 identity switches and 286 track fragmentations. Those counts motivate a closer look at individual trajectories even when the average accuracy is strong.
AMOTA averages recall-normalized tracking accuracy over recall thresholds. AMOTP summarizes position error over recall thresholds. Consult the nuScenes metric definitions for the exact evaluation protocol.
The log’s ID-switch and fragmentation counts are the evaluator’s reported operating-point statistics; they are not averaged in the same way as AMOTA.
03 / Visual diagnosis
The observations are organized below by failure mechanism. Each example distinguishes the recorded observation from the interpretation or improvement it suggests.
The example marks 13 frames of missed pedestrian detection. With no matching measurement, the tracker must balance survival against the chance that the object has left.
Implication: examine existence decay and track management across the gap, as well as the quality of the detections.
The analysis associates high birth rates with premature termination and restart. Lowering the rate to 0.001 fixes several selected examples, including this pair.
Trade-off: making new tracks less likely can also encourage incorrect associations to old tracks.
Several examples terminate during a curved trajectory. One caption explicitly notes that the problem remains after radar velocity is added.
Hypothesis to test: a motion model that better represents turns may help where changing only the birth rate does not.
A car may be detected as a truck, or alternate between other categories. The repository’s class-specific runner makes these errors relevant to association and track continuity.
Direction in the analysis: use classification information over time. The nuScenes output still requires a consistent class label within a track.
The analysis documents identity changes when tracks intersect, including an example of two bus tracks that terminate and restart.
Implication: inspect the competing associations around the intersection, including tracks that temporarily have no detection.
These two values are reported in the archived analysis; the corresponding complete configurations are not supplied there.
Select an example to play its archived animation at 2× speed in a continuous loop.
04 / Method comparison
The source compares outputs labeled MHT and TO-PMB (track-oriented Poisson multi-Bernoulli). These paired visual examples describe individual behavior; they do not establish an overall ranking.
The analysis identifies early termination and sudden position changes in selected MHT outputs. Inspect which association history is retained when several explanations compete.
A separate group records track coalescence and associations made over large distances. Preserving a track is useful only when the identity and trajectory remain credible.
Method labels are preserved from the archived observations. Exact code-to-figure configurations are not available for every pair. Browse all comparison figures.
05 / Lessons and next experiments
The archive suggests several directions. They are presented as research questions, not as demonstrated improvements.
06 / Evidence library
73 image entries from the research archive, including 5 animations. Original captions are retained. GIFs play at 2× speed and loop continuously; repeated figures remain separate entries.
Open a category to inspect its figures. Select any image to view it at full resolution. Some source captions contain only a track token; these are retained without adding an unsupported diagnosis.
07 / Code and provenance
The code spans simulation, nuScenes runners, tracker implementations, and evaluation. Start with the part that matches the question you want to investigate.