A robot observes the same workspace through three synchronized cameras: one wrist camera and two external cameras. An upstream perception system proposes one object track in each view, but one camera may have locked onto a distractor. Determine whether all three tracks describe the same physical object. If they do not, identify the inconsistent camera.
Each row represents one synchronized three-camera episode. The supplied features summarize appearance, motion, timing, and cross-view geometry that would normally be produced by pretrained vision and tracking models.
The target has four classes:
- same: all three tracks are consistent;
- wrist_wrong: the wrist-camera track is inconsistent;
- ext1_wrong: external camera 1 is inconsistent;
- ext2_wrong: external camera 2 is inconsistent.
The test data includes shifts in noise, occlusion, motion, and camera characteristics. Feature scales and missing-value patterns may therefore differ from the training set.
Submit one probability distribution for every episode_id:
episode_id,same,wrist_wrong,ext1_wrong,ext2_wrong
Probabilities must be finite, non-negative, and sum to one in every row.
Submissions are scored using multiclass logarithmic loss:
-mean(log(probability assigned to the true class))
Lower is better. The metric rewards both correct classification and calibrated uncertainty; a confident incorrect prediction is penalized heavily.
The features are precomputed so that the task can be solved quickly without a large GPU. Strong solutions should examine cross-view agreement, handle missing evidence, validate under grouped or shifted splits, and calibrate their output probabilities.
This is an unofficial IOAI-style practice task created by LibrAI.