APRILTAG JPEG BENCHMARK METHODS Run and scope The article uses the archived expanded benchmark on Apple Silicon (macOS 26.5.2, Python 3.12.13). The main study used NumPy 2.5.4, OpenCV 4.14.0.94, Pillow 13.0.0 with libjpeg-turbo 3.2.0, and the pupil-apriltags 1.0.4.post11 detector. The full archive contains 1,244,688 detector evaluations across multiple test suites and settings. The article's full-resolution held-out grid contains 7,200 distinct rendered scenes evaluated at 14 encodings, or 100,800 detector calls. The main study renders one tag36h11 marker with a one-cell white margin into each 640 × 480 RGB frame. Nominal size is the width of the black-bordered square at zero yaw and pitch. The size grid is 10, 12, 16, 20, 24, 32, 48, 64, 96, and 128 px. Yaw is 0, 20, 40, 55, 65, or 75 degrees. Per-parent draws vary pitch within ±15 degrees, roll, subpixel placement, black and white intensity, mild Gaussian blur (sigma 0.4 to 0.7 px), and noise (sigma 0.5 to 2 intensity units). The renderer supersamples the tag region at 3× before area downsampling. The main random seed is 910202611. Each background collection contributes 120 parent groups. A parent is one procedural seed or one source photograph from BSDS500 or COCO128. Each parent contributes every size, yaw, and encoding. Parent indices with remainder 2 modulo 3 are held out, giving 40 parents per collection for evaluation and 80 for development. The same source photo stays in one split. The photographic sources were already JPEG files. The lossless reference preserves the pixels of the rendered composite. Encoding and scoring Each rendered frame goes independently to PNG (compression level 6) and JPEG (4:2:0 chroma subsampling, optimize=False, progressive=False). The tested JPEG settings are 1, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 95, and 100. OpenCV converts the decoded RGB frames to grayscale before detection. The detector uses quad_decimate=1, quad_sigma=0, refine_edges=1, decode_sharpening=0.25, one thread, and the binding's limit of two corrected bits. For each candidate, subtract half a pixel from the native corner coordinates to match the renderer, then align the corners by cyclic order and winding to minimize their root mean square error (RMSE) against the projected black-border corners. The closest candidate matches the tag's location if its RMSE is at most the greater of 3 px or 20% of the mean projected side length. A matched detection counts as correct when its ID also matches. Recall is the fraction of scenes with a correct detection. The corner error percentile uses correct detections. File size includes headers. One KB is 1,000 bytes. The development search pools all 6,000 scenes with nominal size at least 32 px and yaw no greater than 65 degrees. Choose the lowest JPEG setting at which it and every higher tested setting have at least 99% recall, lose at most 1% of PNG-detectable tags, and keep p95 corner RMSE at or below 0.5 px. Evaluate the chosen setting once on the 3,000 held-out scenes in the same range. The later sweep at 32 px and 65 degrees is descriptive and uses the same held-out scenes that revealed the misses. The primary contrast compares the PNG-to-Q20 loss in recall for 12, 16, 20, and 24 px tags with the loss for 64, 96, and 128 px tags. Its 95% interval uses the 2.5th and 97.5th percentiles of 4,000 bootstrap draws of whole parents, stratified by background collection. Data files heldout-size-angle.csv contains the 840 pooled size, yaw, and encoding cells for the held-out full-resolution detector. nominal_size is in pixels and yaw is in degrees. heldout-pooled.csv combines these cells by encoding. heldout-32px-65deg.csv gives the full sweep at the view discussed in the article. In those tables quality 0 denotes PNG, n is the number of scenes, and correct and png_correct are the correct-ID counts for the row's encoding and its PNG baseline. lost counts PNG successes missed in JPEG and gained counts JPEG successes missed in PNG. recall_pct = 100 × correct / n and mean_bytes is per frame. development-selection.csv records the pooled screening results. lost_fraction_png divides lost by png_correct. corner_rmse_p95 is in pixels. primary-contrast.csv gives the contrast in percentage points and counts parent groups in the clusters column. audit.json records the archived macOS audit, including the consolidated measurements and benchmark script SHA-256 digests. The build of the original preview checked the 840 plotted cells against 100,800 archived held-out detector rows. The consolidated archive, benchmark source, and dataset photographs are held separately from these downloads. Pilot illustration The cropped tag images come from the pilot scene positive_0001_024_50, tag36h11 ID 415, nominal 24 px at 50 degrees yaw. The original figure builder checked all four encoded files against their archived SHA-256 digests before taking 40 × 40 crops. PNG, Q70, and Q20 decoded the correct ID. Q5 missed the tag. The pilot was separate from the expanded held-out sweep.