Monocular depth discrimination in natural scenes: Humans vs. deep networks
Bibliographic record
Abstract
Objective. Humans use a number of monocular cues to estimate depth, but little is known about how accuracy varies with depth, nor how we compare to recent deep network models for monocular depth estimation. Here we measure and compare monocular depth acuity for humans and deep network models. Methods. Stimuli were drawn from natural outdoor scenes of the SYNS database of spherical imagery with registered ground truth range data. From each spherical image we extracted 62×49 deg sub-images sampled at regular intervals along the horizon. Four observers viewed randomly-selected images monocularly. Two points on the image were indicated by coloured crosshairs; observers were asked to judge which was closer. The difference in depth was varied to sweep out psychometric functions at four mean depths. Four deep network models were run on the same task. Results. Absolute JNDs were found to increase with mean depth faster than a Weber law for humans and most models, possibly due to the increased foreshortening of the ground surface with depth. While humans outperformed deep network models, a kernel regression model that uses only the elevation angle (height in the image) outperformed both for nearer depths, and we found that both humans and the networks struggled when the two points were fixed to have the same elevation. This suggests that both humans and deep networks may rely largely upon this simple elevation cue, although superior human performance at greater depths indicates that humans can recruit additional image cues. While luminance, colour and spatial frequency cues were all correlated with depth, most of the variance is shared with elevation and adding these cues to the kernel regression model failed to improve its performance. Conclusions. While human monocular depth acuity surpasses current state-of-the-art deep networks, both appear to rely heavily upon gaze elevation to estimate depth.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.004 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.000 | 0.001 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".