Add files using upload-large-folder tool

CLASSES.txt

@@ -0,0 +1,7 @@
+0 person
+1 bicycle
+2 car
+3 motorcycle
+4 bus
+5 truck
+6 van

README.md

@@ -0,0 +1,150 @@
+---
+license: other
+task_categories:
+  - object-detection
+  - image-segmentation
+tags:
+  - traffic
+  - aerial
+  - drone
+  - roadside
+  - multi-view
+  - coco
+  - instance-segmentation
+  - autolabel
+size_categories:
+  - 10K<n<100K
+---
+
+# DLR v1 — 2D detection + instance segmentation
+
+Multi-view roadside + drone traffic imagery from three German sites, with 2D
+boxes, per-instance masks and class labels in COCO format.
+
+Each **position** is a time-synchronized pair: a **drone** bird's-eye view
+(3840×2160, 25 Hz) and one or more fixed **Axis** roadside cameras (1920×960) of
+the same intersection.
+
+> **These labels are machine-generated, not human-verified.** They come from a
+> SAM 3 text-prompted autolabel pipeline plus an automatic postprocessing pass
+> (shadow/reflection repair, de-flickering, track re-association, class voting).
+> Expect autolabel-grade noise. See *Provenance* and *Limitations*.
+
+## Contents
+
+```
+train/  _annotations.coco.json  +  18015 .jpg
+valid/  _annotations.coco.json  +   5892 .jpg
+test/   _annotations.coco.json  +   5880 .jpg
+```
+
+| split | images | instances | source runs |
+|---|---|---|---|
+| train | 18,015 | 375,926 | 34 |
+| valid | 5,892 | 104,024 | 24 |
+| test | 5,880 | 105,952 | 24 |
+| **total** | **29,787** | **585,902** | 58 |
+
+### Classes
+
+| id | name | instances |
+|---|---|---|
+| 0 | person | 0 |
+| 1 | bicycle | 0 |
+| 2 | car | 528,354 |
+| 3 | motorcycle | 16,048 |
+| 4 | bus | 3,202 |
+| 5 | truck | 12,549 |
+| 6 | van | 25,749 |
+
+`person` and `bicycle` are kept in the taxonomy for COCO id compatibility but
+carry **no annotations** — the v1 prompt set targets vehicles only. `van` is
+appended as id 6 so the standard COCO ids 0–5 stay stable.
+
+Traffic is overwhelmingly cars (90%); `bus` is rare (0.5%). Treat this as a
+long-tailed vehicle dataset, not a balanced one.
+
+## Format
+
+Standard COCO. Detection and instance segmentation share one file — masks are
+compressed RLE in `segmentation`, so segmentation is a switch, not a separate
+annotation pass.
+
+```jsonc
+{
+  "id": 1, "image_id": 0, "category_id": 2,
+  "bbox": [x, y, w, h],          // xywh, pixels
+  "area": 3896,                   // == mask pixel count (RLE, not bbox area)
+  "iscrowd": 0,
+  "segmentation": {"size": [h, w], "counts": "..."},   // full-frame RLE
+  "score": 0.96,                  // autolabel confidence
+  "track_id": 1474,               // stable within its source run
+  "global_id": "sb_pos3_drone_0001_1474"   // clip-scoped identity
+}
+```
+
+Loads with `pycocotools`; RLE round-trips (`area == mask.sum()` verified on
+every split).
+
+`global_id` here is **clip-scoped** (`"<run>_<track_id>"`) — it identifies a
+vehicle within one camera, not across cameras. Cross-camera identity comes from
+the drone↔axis merge and ships with the 2.5D localization / Re-ID parts of the
+release, not with this one.
+
+File names are `<run_key>_<source_frame_index>.jpg`, so every image is traceable
+to its source video and frame.
+
+## Splits
+
+Split **by position**, then a frame-level cut — not a random shuffle. Random
+splitting of video frames leaks near-identical neighbours across splits and
+inflates scores.
+
+- **train** positions are exclusive: their frames never appear in valid/test.
+- **eval** positions feed both valid and test, cut temporally (first part →
+  valid, remainder → test). valid and test may share a position; neither shares
+  one with train.
+
+Frames are selected with a gap + min-vehicle rule rather than a fixed stride:
+keep a frame only if it has ≥3 annotations and is ≥25 source frames (~1 s) after
+the last kept one. This drops near-duplicate consecutive frames.
+
+Verified: **0 source-run overlap between train and eval**.
+
+### Known caveat — Tostmannplatz
+
+At the Tostmannplatz site the split is per *camera*: `tp_pos1`'s drone and
+`cam_3_100` are in eval while its `cam_48_101`/`cam_48_106` are in train. Those
+cameras observe **the same intersection at the same instant** from different
+viewpoints, so the same traffic appears on both sides of the split, seen from
+different angles. This is inherited from the v0.1 split policy and kept
+deliberately so v1 numbers stay comparable to v0.1. If you need a strictly
+scene-disjoint evaluation, exclude the `tp_pos1_*` runs from train.
+
+## Provenance
+
+- **Autolabel**: SAM 3 image-mode, per-frame text-prompted grounding with tiling
+  and class-aware NMS; mask-IoU tracker for `track_id`.
+- **Postprocess**: drone cast-shadow trimming, wet-road reflection suppression,
+  de-flickering, global track re-association, score-weighted class voting.
+  Detections the postprocessor rejected are excluded here.
+- **Axis labels** additionally pass through a drone-guided cross-camera bridging
+  stage, which stitches axis tracks that occlusions had fragmented.
+- No LLM is involved in class assignment; classes come from the detector plus a
+  score-weighted per-track majority vote.
+
+## Limitations
+
+- Machine-generated labels; no human verification pass. Small/distant vehicles
+  and heavy occlusions are the weakest cases.
+- Strongly class-imbalanced (90% car).
+- `person`/`bicycle` are empty by construction.
+- Frames within a split are temporally correlated (~1 s apart within a run).
+- Site/weather diversity is limited: three sites, a handful of sessions.
+
+## Citation
+
+Paper in preparation. Please contact the author before using this in a
+publication.
+
+Contact: Javi Borau — `jborau@caos.uc3m.es`

splits.json

@@ -0,0 +1,119 @@
+{
+ "version": "v1",
+ "protocol": "split-by-position-group, then frame-level eval cut",
+ "policy": "TRAIN positions are exclusive (their frames appear only in train). EVAL positions feed BOTH valid and test: temporally-ordered kept frames are cut at val_test_ratio (first part -> valid, rest -> test). train never shares a position with valid/test.",
+ "val_test_ratio": 0.5,
+ "frame_selection": {
+  "rule": "gap + min-vehicles",
+  "min_gap_frames": 25,
+  "min_objects": 3
+ },
+ "caveat_tostmannplatz": "tp_pos1 is split per CAMERA: drone + cam_3_100 are eval while cam_48_101/106 are train. Those cameras see the same intersection at the same instant from different viewpoints. Inherited from v0.1 and kept for comparability; exclude tp_pos1_* from train for a strictly scene-disjoint eval.",
+ "counts": {
+  "train": {
+   "images": 18015,
+   "instances": 375926,
+   "runs": 34
+  },
+  "valid": {
+   "images": 5892,
+   "instances": 104024,
+   "runs": 24
+  },
+  "test": {
+   "images": 5880,
+   "instances": 105952,
+   "runs": 24
+  }
+ },
+ "runs_per_split": {
+  "train": [
+   "kv_pos1_axis_cam3",
+   "kv_pos1_drone_0015",
+   "kv_pos1_drone_0016",
+   "kv_pos1_drone_0017",
+   "kv_pos1_drone_0018",
+   "sb_pos1_axis_cam3",
+   "sb_pos1_axis_cam6",
+   "sb_pos1_drone_0010",
+   "sb_pos1_drone_0011",
+   "sb_pos1_drone_0012",
+   "sb_pos1_drone_0013",
+   "sb_pos4_s1_axis_cam4",
+   "sb_pos4_s1_drone_0031",
+   "sb_pos4_s1_drone_0032",
+   "sb_pos4_s1_drone_0033",
+   "sb_pos4_s2_axis_cam4",
+   "sb_pos4_s2_drone_0036",
+   "sb_pos4_s2_drone_0037",
+   "sb_pos4_s3_axis_cam4",
+   "sb_pos4_s3_drone_0038",
+   "sb_pos4_s3_drone_0039",
+   "sb_pos4_s3_drone_0040",
+   "sb_pos4_s3_drone_0041",
+   "sb_pos4_s3_drone_0042",
+   "sb_pos4_s3_drone_0043",
+   "tp_f1_axis_cam48_101",
+   "tp_f1_axis_cam48_106",
+   "tp_f2_axis_cam48_106",
+   "tp_rieke_axis_cam3",
+   "tp_rieke_axis_cam48_101",
+   "tp_rieke_drone_0019",
+   "tp_rieke_drone_0020",
+   "tp_rieke_drone_0021",
+   "tp_rieke_drone_0022"
+  ],
+  "valid": [
+   "kv_pos2_axis_cam3",
+   "kv_pos2_drone_0023",
+   "kv_pos2_drone_0024",
+   "kv_pos2_drone_0025",
+   "kv_pos2_drone_0026",
+   "sb_pos2_axis_cam3",
+   "sb_pos2_drone_0006",
+   "sb_pos2_drone_0007",
+   "sb_pos2_drone_0008",
+   "sb_pos2_drone_0009",
+   "sb_pos3_axis_cam3",
+   "sb_pos3_drone_0001",
+   "sb_pos3_drone_0002",
+   "sb_pos3_drone_0003",
+   "sb_pos3_drone_0004",
+   "tp_f1_axis_cam3",
+   "tp_f1_drone_0019",
+   "tp_f1_drone_0020",
+   "tp_f1_drone_0021",
+   "tp_f1_drone_0022",
+   "tp_f2_axis_cam3",
+   "tp_f2_drone_0015",
+   "tp_f2_drone_0017",
+   "tp_f2_drone_0018"
+  ],
+  "test": [
+   "kv_pos2_axis_cam3",
+   "kv_pos2_drone_0023",
+   "kv_pos2_drone_0024",
+   "kv_pos2_drone_0025",
+   "kv_pos2_drone_0026",
+   "sb_pos2_axis_cam3",
+   "sb_pos2_drone_0006",
+   "sb_pos2_drone_0007",
+   "sb_pos2_drone_0008",
+   "sb_pos2_drone_0009",
+   "sb_pos3_axis_cam3",
+   "sb_pos3_drone_0001",
+   "sb_pos3_drone_0002",
+   "sb_pos3_drone_0003",
+   "sb_pos3_drone_0004",
+   "tp_f1_axis_cam3",
+   "tp_f1_drone_0019",
+   "tp_f1_drone_0020",
+   "tp_f1_drone_0021",
+   "tp_f1_drone_0022",
+   "tp_f2_axis_cam3",
+   "tp_f2_drone_0015",
+   "tp_f2_drone_0017",
+   "tp_f2_drone_0018"
+  ]
+ }
+}

taxonomy.json

@@ -0,0 +1,51 @@
+{
+ "name": "DLR v1 taxonomy",
+ "version": "v1",
+ "note": "Canonical COCO-order ids 0-5 plus 'van' appended as id 6 so the COCO ids stay stable. Source SAM3 ids: 1=car 2=truck 3=bus 4=motorbike 5=bike 6=pedestrian 7=van. The v1 prompt set is vehicles-only, so person(0) and bicycle(1) carry no annotations.",
+ "categories": [
+  {
+   "id": 0,
+   "name": "person",
+   "instances": 0
+  },
+  {
+   "id": 1,
+   "name": "bicycle",
+   "instances": 0
+  },
+  {
+   "id": 2,
+   "name": "car",
+   "instances": 528354
+  },
+  {
+   "id": 3,
+   "name": "motorcycle",
+   "instances": 16048
+  },
+  {
+   "id": 4,
+   "name": "bus",
+   "instances": 3202
+  },
+  {
+   "id": 5,
+   "name": "truck",
+   "instances": 12549
+  },
+  {
+   "id": 6,
+   "name": "van",
+   "instances": 25749
+  }
+ ],
+ "sam_to_canonical": {
+  "1": 2,
+  "2": 5,
+  "3": 4,
+  "4": 3,
+  "5": 1,
+  "6": 0,
+  "7": 6
+ }
+}