README.md

Figures

Evidence figures kept in git for the research paper.

Training and validation output normally lands in runs/, which .gitignore excludes because Ultralytics writes hundreds of megabytes of checkpoints and batch previews there. The small subset that is actual paper evidence is copied here instead, so it survives independently of one laptop's runs/ folder.

When a new training run replaces the current detector, re-copy the same files and update the numbers below in the same commit, so the figures and the claims never drift apart.

dumbbell_detector/

Dumbbell/weight object detector, run dumbbell_combined_yolo26n, trained from the dumbbell_detection profile in configs/ultralytics_training_config.yaml.

Dataset — data/datasets/dumbbell_combined_yolo26, merged from the two Roboflow dumbbell sets. Classes: dumbbell, weight, other.

Split	Images
train	5668
valid	1093
test	571

Final metrics (epoch 80, from results.csv):

Metric	Value
precision (B)	0.919
recall (B)	0.869
mAP50 (B)	0.919
mAP50-95 (B)	0.744

Per-class behaviour, derived from the raw counts in confusion_matrix.png:

Class	Instances	Predictions	Missed	Precision	Recall
dumbbell	2,153	2,091	361	0.857	0.832
weight	512	518	19	0.952	0.963
other	977	919	153	0.897	0.843

How to read the background column

The normalized matrix is column-scaled by true class, so each column sums to

1. The background column therefore shows how the detector's unmatched detections distribute across classes — not a false-positive rate over background regions.

There were 403 detections that matched no annotation: 289 dumbbell (0.72), 91 other (0.23), 23 weight (0.06). The 0.72 means "72% of the false positives were dumbbell predictions". It does not mean "72% of background regions are detected as dumbbells" — that statement would need a denominator this evaluation does not define, which requires a negative set or dumbbell-free video.

For dumbbell specifically: 289 of 2,091 predictions (13.8%) matched nothing, while 361 of 2,153 annotated instances (16.8%) were missed. The detector fails slightly more by omission than by spurious detection.

Threat to validity: split contamination

A SHA-1 audit of all 7,332 images found 7,276 unique digests: 49 duplicate groups, 24 of which span splits. 24 evaluation images are byte-identical to a training image (20 of 1,093 validation, 4 of 571 test).

More significantly, grouping filenames by source prefix shows 5,669 of 7,332 images (77.3%) belong to source groups whose frames appear in more than one split — these are consecutive frames from continuous video, split randomly. Evaluation images are therefore frequently near-duplicates of training images.

Treat these metrics as an upper bound, not as an estimate of generalization. They document the model deployed in the system. Any generalization claim, or any comparison against another detector, needs a source-disjoint re-partition and a retrain first.

Training configuration

yolo26n.pt, 80 epochs, 640 px, single GPU, seed 0, deterministic, ~3 h 05 min. Note that the dumbbell_detection profile in configs/ultralytics_training_config.yaml now specifies 120 epochs; this archived run predates that value. args.yaml records what was actually used.

File	Contents
results.csv	Per-epoch metrics. The source of every number above.
results.png	Loss and metric curves across all 80 epochs.
confusion_matrix.png	Raw counts per class.
confusion_matrix_normalized.png	Column-normalized version; the one to cite.
BoxPR_curve.png	Precision-recall curve per class.
BoxF1_curve.png	F1 against confidence threshold; useful for justifying the runtime's confidence floors.
args.yaml	The exact Ultralytics arguments this run used.