Access is granted for non-commercial research and evaluation use only. By requesting access you agree not to redistribute this dataset or any derivative data, to cite Dhi Technologies in any publication or output that uses it, and to obtain a separate commercial license via dhi-tech.com before any commercial use. Access requests are reviewed manually by Dhi Technologies.

Amodal Counting Benchmark

Product: amodal-counting, "count what detectors can't see": visibility-corrected crowd counting with calibrated count intervals.

Paper: Dhi Labs paper 07, Inference-Aligned Detectability Calibration and Conformal Count Intervals (manuscript in preparation for Computer Vision and Image Understanding). Every number on this card is taken from that manuscript.

What this dataset is

Procedurally generated scenes with exact ground truth: a simulated detector with a known detectability curve, the true count and true visibility of every object, and the detector's reported boxes. It ships so that the scoring can be re-run on the same files. This dataset contains no real-world imagery.

What's in this dataset

scenes.jsonl (6,000 rows, about 6.8 MB): 5 occlusion-density levels (crowding in {0.0, 0.2, 0.4, 0.6, 0.8}) by 3 generation seeds (seed_run in {10, 11, 12}), 400 scenes per pair. Within each block, the first 240 rows in generation order have "split": "train" and the last 160 have "split": "eval". Each row:
```
{"crowding": 0.0, "image_size": [640, 480],
 "detections": [[x1, y1, x2, y2], ...],
 "occluders": [[x1, y1, x2, y2], ...],
 "true_count": 12, "detected_count": 10,
 "true_visibilities": [1.0, 1.0, ...], "seed": 1000000,
 "seed_run": 10, "split": "train"}
```
detections are the simulated detector's reported boxes. true_count and true_visibilities are ground truth, not predictions.
bench_results.json (about 8.4 KB): the output of three runs of the earlier five-level benchmark, one per seed_run, with a per_seed list and an aggregate block.

Scope of these files. These files come from an earlier release with five crowding levels and three generation seeds. The results below come from the paper 07 runs (nine crowding levels, five fitted seeds, 4,500 test scenes). They are not recomputed from these files.

Results (paper 07)

Synthetic generator, nine crowding levels, five fitted seeds, 4,500 test scenes

The aligned fit has lower MAE than the true-visibility fit at every crowding level, and MSE falls at all nine levels.
Coverage at nominal 0.90: unscaled analytic intervals cover 0.546 to 0.640. Conformal width calibrated on disjoint scenes covers 0.876 to 0.920.
The estimation variance of the correction is 0.18 to 1.89 percent of the mean squared error. Its delta-method predictor lies within a factor of 1.2 of the simulated value.

Public crowd-counting benchmark (Part A and Part B)

Estimation variance is 0.07 and 0.35 percent of the mean squared error on Parts A and B.
A frozen point detector trained on Part A, with supervised calibration on 400 Part B training images, lowers MAE from 22.17 to 10.71 on all 316 Part B test images. The paired improvement is 11.46, with a 95 percent interval of 9.14 to 13.94.
Part A conformal intervals reach their marginal coverage. The count-relative score brings the densest crowd tercile to 0.903 to 0.949.
The estimator, its improvement condition and its conformal intervals apply after any point or box detector.

Bias-variance condition. An exact bias-variance decomposition, which keeps scene-level covariance, gives the condition under which a corrected count has lower mean squared error.

How the method works

Visibility from box geometry. Each detection's visibility is the fraction of its box not covered by occluders, other detections in front of it, or the image border.
Inference-aligned detectability calibration. A logistic detectability curve is fit on visibility computed the way the deployed pipeline computes it. Queries outside the calibrated support are flagged.
Horvitz-Thompson counting. Each detection is weighted by the inverse of its detection probability, and the weights sum to the corrected count.
Conformal interval calibration. A split-conformal multiplier on the analytic width makes the stated coverage empirical.

Limitations

The synthetic results hold for the generator's stated assumptions. The public-benchmark results hold for the detector, splits and calibration described in paper 07.
Real-world validation beyond these benchmarks is in progress.

Schema notes

Boxes are [x1, y1, x2, y2] in pixels, with the image origin at the top left.
true_visibilities is the ground-truth fraction of each true object that is not covered by occluders, other objects or the frame border.
seed_run identifies the generation seed of a row. seed is the per-scene seed (seed_run * 100_000 + row_index_in_block).
split ("train" or "eval") marks the 240/160 partition of each (crowding, seed_run) block.

How to load it

import json
from huggingface_hub import hf_hub_download

repo_id = "Dhi-Technologies/amodal-counting-benchmark"
scenes_path = hf_hub_download(repo_id, "scenes.jsonl", repo_type="dataset")
bench_path = hf_hub_download(repo_id, "bench_results.json", repo_type="dataset")

scenes = [json.loads(line) for line in open(scenes_path)]
bench = json.load(open(bench_path))

print(len(scenes), "scenes")
print(scenes[0]["crowding"], scenes[0]["true_count"], scenes[0]["detected_count"],
      scenes[0]["seed_run"], scenes[0]["split"])

No datasets loader script ships here. The files are JSON Lines and JSON, readable with the standard library once downloaded.

License

This dataset is released under CC BY-NC 4.0 (non-commercial). Access is gated and requires manual approval. It is provided for non-commercial research and evaluation only. Redistribution is not permitted, and any publication or output using it should cite Dhi Technologies. Commercial use requires a separate agreement; contact dhi-tech.com.