Asymmetric Degradation in Deep Stereo: Accuracy and Left-Right Consistency

Stereo rigs in the field often degrade on one side only: one lens defocuses or catches a flare, one sensor becomes noisy. We run a controlled comparison of left-only, right-only, and both-view degradation at matched severity on the original Spring training data, using seeded defocus, flare, and sensor noise at three severities, for two public SceneFlow checkpoints (RAFT-Stereo, PSMNet) on all 37 sequences. The effect depends on the corruption. For defocus, blurring one view is far worse than blurring both: at the highest severity the median RAFT-Stereo error is 1.06 px with both views blurred but 16.64 and 284.73 px with only the left or right view blurred (clean: 0.54 px). Flare shows the same pattern for PSMNet only. For noise, one-sided degradation was not significantly worse, and for PSMNet two-sided noise was significantly worse. At the standard 1-px threshold, the left–right (LR) check rejects most of the one-sided defocus output: coverage falls from 89.3% to 46.2% (RAFT-Stereo), and fewer wrong pixels survive. The pixels it keeps, however, are more often wrong (retained error 3.8% vs. 16.3%), and the LR check’s ranking quality drops. The LR check is therefore not a sufficient safeguard in either mode. The feature-correlation entropy we tested was a weaker detector than LR consistency in every setting. Code and the raw per-frame log are released.