How it works
Reading the results
What a map does and does not say, and what to report with it.
Three things, kept apart
| Comes from | Says | |
|---|---|---|
| Attribution | Grad-CAM | where, at one layer, the score’s gradient weights active features |
| Perturbation effect | occlusion | how the score changes when a region is replaced by a baseline |
| Biological interpretation | — | which body part or behaviour matters to the animal |
YORU-Explain produces the first two and never the third. Both are statements about one model’s score for one detection. A claim about body parts needs independently measured features and controls; named regions such as wings, head or legs, with matched-area random controls, are planned (Roadmap).
Do not tune until it looks right
Choose the settings before looking at the maps, and report the maps as they come out. A map that does not highlight the wings or the legs is a measurement of the model, not a failure of the method — and adjusting the patch size or the baseline until it does would make the figure say what you expected rather than what the model does.
Settings change the numbers
- Baseline, patch size, stride. On the copulation event of YORU’s test video, the largest score change was 0.055 with hot-pink patches and 0.523 with black ones (Baselines). Name the baseline in every figure, and show more than one before making a claim.
- Identity mode. Fixed-candidate and re-detection answer different questions and give different per-patch numbers (Identity modes).
- Batch size and device. Batch size 1 reproduces YORU exactly; larger batches have a noise floor of about 2.7 × 10−3 (Batch size). CUDA (fp16 autocast, as YORU runs) and the CPU (fp32) give slightly different scores, e.g. 0.857316 and 0.857245 for the same detection.
Frames are not independent samples
Consecutive frames of one video, one animal, one day are not independent replicates. The Statistics tab is descriptive and runs no test. A group analysis that treats videos, animals and days as the units is planned.
What to report with a figure
Everything below is written to every run’s config.json and manifest.json:
- the method, the score type and the identity mode;
- for occlusion: the baseline, the patch size, the stride, the region and the batch size;
- for Grad-CAM: the layer and the precision;
- the device, the model file’s SHA-256, and the versions of YORU-Explain and YORU;
- the identity check against YORU.
About the figures on this site
The examples come from YORU’s test video (test_video_fly_copulation.mp4) and a YOLOv5 model trained with YORU’s training script on 900 frames of that same video, as a test fixture for YORU-Explain. They describe that fixture model’s score. They are not findings about flies.