YORU-Explain Why a YORU detection scores as it does

User guide

Statistics and Settings

The per-event numbers of each run, and the settings of the application itself.

Statistics

The Statistics tab showing an occlusion run of three events
The Statistics tab with the occlusion run of three events: the run's summary and settings at the top, one row per event below.

Choose a Run. The summary above the table names the method, the status, the model (path, SHA-256, device), the identity check against YORU, the events by status, the error codes, and every setting of the run. Refresh reads the project’s runs again; Open run folder opens the folder in Explorer.

The table has one row per event:

Column Meaning
event, frame, class the event
status, error ok, or failed with an error code (Output files)
score type, identity what was explained, and how the detection was followed
original the detection’s score on the unchanged frame
largest drop, smallest the largest and the smallest per-patch score change (occlusion)
in box, outside the map’s mean inside and outside the detection’s box (occlusion)
counted patches that entered the map
disappeared, class changed re-detection: patches after which the detection was gone, or only another class matched
candidate dropped fixed-candidate: patches after which NMS would no longer keep the candidate
noop the score change of the unchanged frame run through the same pipeline — a control, expected to be exactly 0
CAM in box Grad-CAM: the share of the map’s mass inside the box

Click a header to sort; double-click a row, or press View in Explore, to see the map.

Descriptive, per event

Frames from one video are not independent samples. No statistical test is computed here, and none should be read into these numbers. They describe the model’s score, not the animal (Reading the results).

Settings

The Settings tab
The Settings tab. The text box below lists versions, hardware, the adapters and — once a model is loaded — everything the adapter knows about it.
Setting  
Compute device auto, cuda or cpu, used when a model is loaded. auto follows YORU: $YORU_DEVICE, else CUDA, else CPU. Scores differ slightly between devices (CUDA runs with fp16 autocast, as YORU does), so every run records its device.
Logs The log file’s path, and Open log folder
Analysis settings file (JSON) Export current settings… saves the settings of the Analyze tab; Import settings… loads a saved file — or a run’s config.json — back into it
Versions, adapters and the loaded model YORU-Explain, Python, NumPy, OpenCV, PyTorch, YORU and the other packages, with the git commits of the YORU and YORU-Explain checkouts; the platform and GPU; the adapter of each YORU backend; and the loaded model’s information and capabilities. Copy it into a bug report.
Menu Items
Window Fit window to this screen, Maximize window, Text size, Save layout now, Reset layout to default — YORU’s own window menu
Project New project…, Open project…, Save
Help Open log folder, the version

Closing the window while a run is going asks first; if you choose to quit, the run is cancelled and its folder closed before the window goes.