rajanish4's picture
cleanup: splits_verification.md
934916d verified
|
Raw
History Blame
6.72 kB

Splits Verification Report

This report is the proof artifact for splits.csv in this dataset repo. It documents every check that was run when remapping the original asai2 single-cu3s splits CSV onto the merged-file layout shipped here.

All numbers below are reproducible — run the scripts in D:/hf_upload_staging/:

  • build_index.py — parses .info files into frame_index.csv
  • build_splits.py — joins the asai2 CSV against the index to emit splits.csv
  • build_per_subfolder_coco.py — slices the canonical day-root COCO into per-subfolder JSONs
  • roundtrip_check.py — opens 44 sampled merged cu3s frames and asserts on-disk frame name matches the predicted camera frame number

Inputs (with sha256)

File sha256
frame_index.csv (derived from .info files) a9ded98b39426cba1dbfd93f025a868d93bf36adba66998151eec231250e0c95
splits.csv (emitted) e12b40aaae4d63853b489b23b5178a238ee85412b83ee278b7e4ee4ad954f84f

1. Coverage — every asai2 row maps to exactly one merged-file row

  • asai2 CSV rows: 1136
  • frame_index.csv rows: 1136
  • Keys in index but not in asai2: 0
  • Keys in asai2 but not in index: 0

✅ Coverage exact, no missing or extra rows.

2. Per-day frame counts

Day Subfolders Saved frames (info → index) Canonical day-root COCO images asai2 CSV image_id range
day2 6 384 384 0..383
day3 6 492 492 0..491
day4 3 260 260 0..259
Total 15 1,136 1,136

Matches whitepaper/dataset_summary.json exactly.

3. Per-subfolder coverage (15 / 15 reconciled)

For each subfolder, the saved-frame count in the .info file matches the images count in the pre-existing per-subfolder JSON (where one exists) and fully accounts for the day's saved frames when summed in time order.

Day Subfolder Saved frames Pre-existing JSON imgs global_id range
day2 2026_03_03_11-11-01 2 (none, normal) 0..1
day2 2026_03_03_11-31-31 17 (none, normal) 2..18
day2 2026_03_03_11-38-39 81 (none, normal) 19..99
day2 2026_03_03_13-58-04_1 96 (none, normal) 100..195
day2 2026_03_03_13-58-04_2 136 136 196..331
day2 2026_03_03_15-25-02 52 52 332..383
day3 2026_03_10_10-17-20 84 84 0..83
day3 2026_03_10_10-58-55 36 36 84..119
day3 2026_03_10_11-30-45 120 (none, normal) 120..239
day3 2026_03_10_12-00-18 40 40 240..279
day3 2026_03_10_14-32-01 92 92 280..371
day3 2026_03_10_15-12-17 120 120 372..491
day4 2026_03_17_11-11-50 80 80 0..79
day4 2026_03_17_11-41-54 80 80 80..159
day4 2026_03_17_14-38-58 100 100 160..259

✅ Every pre-existing per-subfolder JSON's image count equals the count of saved frames in the same subfolder's .info file.

Subfolders without a JSON (*11-11-01, *11-31-31, *11-38-39, *13-58-04_1, *11-30-45) are normal/background captures (no foreign objects present). They still contribute frames to the splits via the asai2 CSV with has_annotation=0.

4. Annotation-equivalence — per-subfolder slice = canonical day-root slice

For each subfolder where a pre-existing per-subfolder JSON exists, the regenerated <subfolder>.json (sliced from the canonical day-root COCO with image_ids rebased) agrees on three independent measures:

Day Subfolder Pre-existing anns Generated anns Match Per-class multiset match
day2 13-58-04_2 276 276
day2 15-25-02 92 92
day3 10-17-20 84 84
day3 10-58-55 36 36
day3 12-00-18 80 80
day3 14-32-01 88 88
day3 15-12-17 360 360
day4 11-11-50 260 260
day4 11-41-54 120 120
day4 14-38-58 140 140
Total 1,536 1,536

The 1,536 figure matches whitepaper/dataset_summary.json["object_counts"] exactly.

Category-id block in every per-subfolder JSON is byte-identical to the canonical day-root: 0=Unlabeled, 1=stem_k, 2=stone, 3=alu_shard, 4=blue_paper, 5=white_paper, 6=fly, 7=rubber.

5. Split-preservation — distribution identical to asai2

Split asai2 row count merged splits.csv row count match
train 808 808
val 148 148
test 180 180
Total 1136 1136

Ratio (71.13 / 13.03 / 15.85 %) matches whitepaper §"Splits".

Per-(day, split) annotated-frame counts

train val test
day2 136 24 28
day3 228 52 48
day4 136 8 36
All 500 84 112

(asai2 and merged splits.csv produce identical values for every cell.)

6. No-leakage — no group_id straddles two splits

splits.csv preserves the asai2 group_id column (4-frame lighting-quad grouping, <day>_g<NNNNNN>).

  • Total distinct group_ids: 284
  • Group_ids whose rows span more than one split: 0

✅ No information leakage across the train/val/test boundary at the lighting-quad level.

7. Round-trip — physical-frame check via cuvis SDK

44 frames (≈3 per subfolder: first, middle, last) sampled across all 15 merged cu3s files. For each, opened the cu3s with cuvis.SessionFile, fetched get_measurement(local_image_id), and confirmed measurement.name ends in the same camera frame number that the index predicts.

  • Sampled frames: 44
  • Passes: 44
  • Failures: 0

Sample-of-the-samples:

day / subfolder local_image_id global_image_id expected camera frame mesu.name
day2 / 2026_03_03_15-25-02 0 332 152 Auto_000_0152
day2 / 2026_03_03_15-25-02 26 358 Auto_000_… (✓)
day3 / 2026_03_10_15-12-17 119 491 Auto_000_… (✓)
day4 / 2026_03_17_11-11-50 0 0 1992 Auto_000_1992

(Full record: roundtrip_report.json.)

Result

All seven checks pass without exception. splits.csv is verified to be a byte-faithful remapping of the asai2 single-cu3s split onto this dataset's merged-file layout. No frame is lost, no annotation is moved, and no split is altered.