kitti_eval: name class 3 the class clean_data actually filters for

[egeboy35]

clean_data decides which annotations belong to a class index:

CLASS_NAMES = ['car', 'pedestrian', 'cyclist', 'van', 'person_sitting', 'truck']

get_official_eval_result names the same indices for the results table, and inverts that map so a caller can pass a class name:

3: 'Sign', #'Van',

Index by index, it is the only one of six that disagrees

0: filter=car             label=Car             same
1: filter=pedestrian      label=Pedestrian      same
2: filter=cyclist         label=Cyclist         same
3: filter=van             label=Sign            <-- differs
4: filter=person_sitting  label=Person_sitting  same
5: filter=truck           label=Truck           same

Two consequences:

• name_to_class['Sign'] = 3, so a caller asking for signs is given the van numbers, under the heading "Sign".
• name_to_class['Van'] is absent, so a caller using KITTI's own class name gets a KeyError.

Three things in the repo say which one was meant

1. get_coco_eval_result, 120 lines further down the same file, has 3: 'Van'.
2. clean_data line 50 treats Van as a KITTI class in its own right: current_cls_name == "car" and gt_name == "van" → neighbour, not a false positive.
3. min_overlaps gives index 3 a 0.7 IoU threshold — the vehicle threshold, where pedestrian and cyclist get 0.5.

And Sign is a Waymo class, not a KITTI one. This repository already says so: detection3d/waymo2kitti_async.py:87 is 'SIGN': 'Sign' # not in kitti.

The commented-out #'Van' on the same line is what this restores.

Tests

Adds kitti_eval/tests/test_class_names_agree.py — 10 tests. They read the source with ast rather than importing eval.py, which compiles its kernels with @numba.jit at import time; nothing here makes a claim about numba. Standard library only, no dataset, no GPU.

Against this branch: 10 passed. Against the file as it stands on main: 4 failed, 6 passed:

FAILED test_every_index_names_the_class_clean_data_filters_for[get_official_eval_result]
FAILED test_the_two_maps_agree_with_each_other
FAILED test_a_caller_can_name_every_class_the_filter_accepts[get_official_eval_result]
FAILED test_no_index_map_carries_a_class_kitti_does_not_have

The same two parametrised tests pass for get_coco_eval_result either way — which is the internal contradiction showing up in the test report rather than in prose.

pytest DeepDataMiningLearning/ngperception/detection/kitti_eval/tests