README.md

---
license: other
task_categories:
  - object-detection
  - image-segmentation
tags:
  - traffic
  - aerial
  - drone
  - roadside
  - multi-view
  - coco
  - instance-segmentation
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) and one or more fixed **Axis** roadside cameras (1920×960) watching
the same intersection.

## Contents

```
train/  _annotations.coco.json  +  <run_key>/*.jpg   (34 subdirs, 18015 images)
valid/  _annotations.coco.json  +  *.jpg             (5892 images)
test/   _annotations.coco.json  +  *.jpg             (5880 images)
```

| 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 |

`train/` images are grouped into one subdirectory per source run because the Hub
caps a directory at 10,000 files. This is transparent to loaders: COCO
`file_name` is a relative path, so `os.path.join(split_dir, file_name)` resolves
correctly in all three splits.

## 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 for COCO id compatibility but carry **no
annotations** — the prompt set targets vehicles only. `van` is id 6 so the
standard COCO ids 0–5 stay stable.

Traffic is overwhelmingly cars (90%); `bus` is rare (0.5%). This is 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`.

```jsonc
{
  "id": 1, "image_id": 0, "category_id": 2,
  "bbox": [x, y, w, h],          // xywh, pixels
  "area": 3896,                   // == mask pixel count, not bbox area
  "iscrowd": 0,
  "segmentation": {"size": [h, w], "counts": "..."},   // full-frame RLE
  "score": 0.96,                  // detection confidence
  "track_id": 1474,               // stable within its source run
  "global_id": "sb_pos3_drone_0001_1474"
}
```

Loads with `pycocotools`; RLE round-trips (`area == mask.sum()` verified on
every split).

`global_id` is **clip-scoped** (`"<run>_<track_id>"`) — it identifies a vehicle
within one camera, not across cameras.

Image file names are `<run_key>_<source_frame_index>.jpg`, so every image is
traceable to its source video and frame.

```python
import os
from pycocotools.coco import COCO

split = "train"
coco = COCO(f"{split}/_annotations.coco.json")
img = coco.loadImgs(coco.getImgIds()[0])[0]
path = os.path.join(split, img["file_name"])          # works for every split
anns = coco.loadAnns(coco.getAnnIds(imgIds=img["id"]))
mask = coco.annToMask(anns[0])                        # HxW binary
```

## Splits

Split **by position**, then a frame-level cut — not a random shuffle. Randomly
splitting 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, which drops near-duplicate consecutive frames.

Verified: **0 source-run overlap between train and eval**.

### Caveat — Tostmannplatz

At Tostmannplatz 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. Exclude the `tp_pos1_*` runs from train if you need a strictly
scene-disjoint evaluation.

## Notes

- Strongly class-imbalanced (90% car).
- `person`/`bicycle` are empty by construction.
- Frames within a split are temporally correlated (~1 s apart within a run).
- Three sites only; limited weather and lighting diversity.

Contact: Javi Borau — `jborau@caos.uc3m.es`