NVIDIA Clear Field — visual review

18 clear-weather clips, built to the NFF schema as the clear-weather counterpart to the 15 foggy clips.

Local only. The NVIDIA AV Dataset Licence forbids redistributing the data or derivatives, including the frames on this page. Do not publish or host this file.

clips, 19.9 s each

countries

179,443

3-D boxes

3,501

tracks

19.1 M

radar returns

7 + 7

camera streams per clip

Figures

Generated by _evaluation/bin/make_figures.py and make_overview.py from the CSVs in _evaluation/stats/.

The headline comparison: the four clearest clear clips above the four densest fog clips, same camera and pipeline.

Every clip placed on the two visibility metrics. The sets do not overlap on either one.

The whole clear set, one front-wide frame per clip.

Geographic and lighting spread of the 18 clips.

Capture month — the fog set is Feb–Mar only, so the clear set deliberately spreads.

3-D box counts per class, clear against fog, on the real NCore taxonomy.

How crowded each clip is: tracks against total boxes, sized by person annotations.

Ego motion per clip — the set spans stop-and-go traffic to open motorway.

How long objects stay visible, which bounds what a reconstruction can fuse over time.

Radar volume per clip and which clips carry the extra pair of side radars.

LiDAR beam return fraction. Unlike the camera metrics, this one overlaps between the sets.

Camera-to-LiDAR time offsets, measured across all 18 clips.

Per-clip sheets

Each sheet shows all 7 rectified cameras, the LiDAR and radar bird's-eye views, and the original-fisheye / undistorted pair — every panel at t = 10.0 s. Click a clip to open it.

NCF-01 Austria · day mountain town 74 tracks · 623 person · 8 radar · vis 1,980 m
NCF-01__e9d14055 · source e9d14055 · chunk 1422

NCF-02 Denmark · day urban day 114 tracks · 1,322 person · 8 radar · vis 1,849 m
NCF-02__02805273 · source 02805273 · chunk 2155

NCF-03 France · day Mediterranean palms 188 tracks · 248 person · 8 radar · vis 1,422 m
NCF-03__882e3305 · source 882e3305 · chunk 1382

NCF-04 Germany · day motorway underpass 124 tracks · 534 person · 10 radar · vis 2,032 m
NCF-04__636bccdc · source 636bccdc · chunk 2935

NCF-05 Germany · day open farmland road 216 tracks · 0 person · 10 radar · vis 1,271 m
NCF-05__7c0feb79 · source 7c0feb79 · chunk 134

NCF-06 Italy · day industrial edge 110 tracks · 0 person · 8 radar · vis 1,759 m
NCF-06__1f918b61 · source 1f918b61 · chunk 1392

NCF-07 Poland · day historic classical 113 tracks · 6,673 person · 8 radar · vis 2,034 m
NCF-07__03d21f72 · source 03d21f72 · chunk 288

NCF-08 Poland · day urban square, pedestrians 724 tracks · 19,396 person · 8 radar · vis 2,065 m
NCF-08__3ece764c · source 3ece764c · chunk 169

NCF-09 Portugal · day urban, blue sky 205 tracks · 5,093 person · 8 radar · vis 2,687 m
NCF-09__c4e82838 · source c4e82838 · chunk 751

NCF-10 Romania · day historic town 236 tracks · 9,662 person · 8 radar · vis 2,107 m
NCF-10__a6b616ab · source a6b616ab · chunk 372

NCF-11 Romania · day tree avenue 192 tracks · 3,766 person · 8 radar · vis 1,821 m
NCF-11__14288f7d · source 14288f7d · chunk 2985

NCF-12 Spain · day urban viaduct 269 tracks · 1,705 person · 8 radar · vis 1,852 m
NCF-12__66997e09 · source 66997e09 · chunk 281

NCF-13 Sweden · day residential houses 29 tracks · 220 person · 8 radar · vis 1,711 m
NCF-13__55daec54 · source 55daec54 · chunk 1131

NCF-14 United States · day summer tree tunnel 4 tracks · 75 person · 8 radar · vis 1,879 m
NCF-14__5174d6bb · source 5174d6bb · chunk 2916

NCF-15 United States · dusk urban dawn 119 tracks · 255 person · 10 radar · vis 2,054 m
NCF-15__4cae610d · source 4cae610d · chunk 2087

NCF-16 Belgium · night urban night, bus 256 tracks · 3,515 person · 8 radar · vis 2,787 m
NCF-16__6d5142c0 · source 6d5142c0 · chunk 1898

NCF-17 Italy · night urban night, parked 185 tracks · 468 person · 8 radar · vis 1,729 m
NCF-17__795e76dd · source 795e76dd · chunk 359

NCF-18 United States · night night arterial, heavy traffic 343 tracks · 498 person · 10 radar · vis 1,728 m
NCF-18__959308aa · source 959308aa · chunk 576

Rectification v2 (pilot)

v1 rectified all seven cameras onto one global K (fx=640, 1280×960, 90° FOV), which cost the 30° teles 5.8× angular resolution and left 95% of their canvas empty. v2 gives each camera its own K at its native resolution, resampled from the untouched fisheye. v1 is still in place. See RECTIFY_V2.md.

NCF-02v1 vs v2 whole frames + 1:1 detail crops

NCF-07v1 vs v2 whole frames + 1:1 detail crops

NCF-08v1 vs v2 whole frames + 1:1 detail crops

NCF-10v1 vs v2 whole frames + 1:1 detail crops

NCF-15v1 vs v2 whole frames + 1:1 detail crops

Clip table

id	country	light	scene	tracks	boxes	person	radar	speed km/h	path m	vis m	split
NCF-01	Austria	day	mountain town	74	3,588	623	8	42.2	232	1,980	train
NCF-02	Denmark	day	urban day	114	4,541	1,322	8	33.1	182	1,849	train
NCF-03	France	day	Mediterranean palms	188	7,057	248	8	26.7	147	1,422	val
NCF-04	Germany	day	motorway underpass	124	7,612	534	10	14.7	81	2,032	train
NCF-05	Germany	day	open farmland road	216	8,275	0	10	49.2	271	1,271	train
NCF-06	Italy	day	industrial edge	110	8,422	0	8	27.8	153	1,759	train
NCF-07	Poland	day	historic classical	113	8,900	6,673	8	29.2	161	2,034	train
NCF-08	Poland	day	urban square, pedestrians	724	33,528	19,396	8	18.4	101	2,065	val
NCF-09	Portugal	day	urban, blue sky	205	13,414	5,093	8	17.0	94	2,687	train
NCF-10	Romania	day	historic town	236	12,962	9,662	8	31.5	174	2,107	train
NCF-11	Romania	day	tree avenue	192	12,438	3,766	8	11.6	64	1,821	train
NCF-12	Spain	day	urban viaduct	269	8,425	1,705	8	25.6	141	1,852	train
NCF-13	Sweden	day	residential houses	29	2,898	220	8	18.2	100	1,711	val
NCF-14	United States	day	summer tree tunnel	4	98	75	8	33.7	185	1,879	train
NCF-15	United States	dusk	urban dawn	119	6,097	255	10	39.0	214	2,054	test
NCF-16	Belgium	night	urban night, bus	256	11,674	3,515	8	16.8	92	2,787	train
NCF-17	Italy	night	urban night, parked	185	13,991	468	8	20.4	112	1,729	val
NCF-18	United States	night	night arterial, heavy traffic	343	15,523	498	10	34.2	188	1,728	held_out

Built from _evaluation/stats/*.csv and the clips themselves. Regenerate with _evaluation/bin/make_index.py. Video and directory links are relative — keep this file inside _evaluation/.