Find the failures the simulator misses.
A small target-system budget is spent where a cheap proxy is informative, and support-aware mutual information keeps those labels from collapsing onto one failure mode.
Real KITTI
Failure is the missed-car score on real KITTI: one minus the average overlap between the detector’s boxes and the ground-truth cars. A frame counts as severe when that score is at least 0.45. Coverage is the mean distance between labeled frames in feature space, the same cumulative diversity measure as in the paper. Higher means the labels are spread across more of the scenario space.
- Ours
- Proxy only
- Real only
- MI only
- Random
KITTI
Virtual KITTI clone
- Detected
- Missed
- Extra
Loading the KITTI acquisition…
Fifteen clips step through all 85 real labels. The pictures are frames our method labeled.
Acquisition
- Real system
- Proxy
- Mutual information
Loading the 1D example…
Mutual information
SIMPLER
The same grasp, rolled out in the proxy simulator and the target simulator. The interesting case is a success on the proxy and a miss on the target.