Make BAS evaluation runnable: four defects, none of which allowed a single file to load

[AtomScott]

Companion to #21. BAS evaluation was as non-functional as GS-HOTA, for a different set of reasons: not one of the ten released BAS files could be read by the loader that ships to read them.

1. Wrong top-level key

_parse_bas reads data["annotations"]; the files key the event array actions. KeyError on the first line of every file. Now accepts either.

2. Wrong label case

BAS_LABELS declares Title Case ("High Pass"); the data is UPPER CASE ("HIGH PASS"). Unknown labels raise rather than being skipped, so this fails on the first event of every file once (1) is past.

The label sets are identical — purely casing. Labels are now matched case-insensitively and canonicalised to the documented Title Case spelling. A genuinely unrecognised label still raises; silently dropping events would understate recall.

3. A third period in a different format

117092, 132831 and 132877 each carry a further block of events whose gameTime omits the half prefix ("90:01" rather than "1 - 0:01") and whose position continues past 5,400,000 ms:

match	half 1	half 2	extra block
117092	1077	1058	1007
132831	1279	1192	691
132877	1204	1096	431

2,129 events that crashed the gameTime split with "not enough values to unpack". 117092's metadata does declare EXTRA_FIRST_HALF and EXTRA_SECOND_HALF, so these look like genuine extra time. The period is now inferred from position rather than the events being dropped.

4. position is absolute, not per-half — and the docs say otherwise

Verified on 117093:

half 1:  position       680 .. 2,694,000   gameTime "1 - 0:01"  .. "1 - 44:55"
half 2:  position 2,700,760 .. 5,506,280   gameTime "2 - 45:00" .. "2 - 91:46"

Half 2 does not restart near zero. Both the clock and position are absolute from the start of the match.

docs/format-bas.md and sections/02_dataset.tex both state "a position field giving milliseconds from kickoff of the half indicated in gameTime", and the paper gives the cross-reference formula ⌊position/40⌋. Following either misaligns every half-2 event by 45 minutes.

Event.image_id used t_ms directly, so it placed every half-2 event 67,500 frames late. It now derives from a new Event.t_ms_in_half, documented as approximate: it subtracts a nominal 45-minute half, while real halves run over — half-2 in-half times reach ~48 minutes here, and a few events sit slightly before the nominal boundary, giving small negative values. Frame-exact work should use the per-period frameStart in <match>_tracker_box_metadata.xml.

I have not changed the docs or the paper. Which of the two is authoritative is your call — the data could be relabelled to per-half, or the documentation corrected. tests/test_bas_map_identity.py pins the observed behaviour so that a future data release switching to per-half timing fails loudly rather than silently.

Verification

make bas-map-test:

23,663 events parse across all ten matches      (previously: zero)
mAP@1s = 1.0000   mAP@5s = 1.0000              scoring ground truth against itself
all 12 classes at 1.0 individually

As with GS-HOTA, a perfect-prediction score validates parsing, matching and the AP computation. It does not exercise the partial-match path — that needs real predictions.

Note on the test

The test pins sys.path to its own checkout. Without that it silently exercises whatever working tree the venv's editable install points at — which bit me while writing it, and would bite anyone running tests from a second clone.

🤖 Generated with Claude Code