File size: 3,006 Bytes
1ee0b43
0b9cb09
1ee0b43
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0b9cb09
 
 
1ee0b43
 
 
 
 
 
 
 
 
 
0b9cb09
1ee0b43
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0b9cb09
1ee0b43
0b9cb09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
---
license: mit
task_categories:
- text-to-image
- image-to-text
tags:
- person-retrieval
- text-based-person-search
- aicity-challenge
pretty_name: PAB hard-negative-pair annotations for SCOUT
---

# PAB hard-negative-pair annotations

Supporting data for **SCOUT** (Sim-to-Real Text-Based Person Retrieval by Embedding-Space Prediction
over Frozen Video Features), an ECCV 2026 workshop paper on AI City Challenge Track 4 (Text-Based
Person Re-Identification, Sim2Real). Code: https://github.com/abtraore/SCOUT-ECCV

## What this is

The AI City Challenge Track 4 release of PAB (Pedestrian Anomaly Behavior) strips several fields
from the dataset authors' original CMP annotation format. This is a re-fetch of the **original,
richer** per-image annotation files, joined back in by `data/pab.py`'s hard-negative-pair machinery
(`data.use_hard_negs` / `data.hard_neg_annotation_dir` in the training configs, and
`--hard-neg-annotation-dir` in `scripts/local_eval.py` for the `val_hard`/`train_hard` splits).

Each `train/attr_N.json` is JSONL (one record per line, despite the `.json` extension), one record
per training image, with fields including:

- `image`, `image_id`: the PAB image path and id
- `caption`: the training caption for that image
- `hard_i`, `hard_i_id`: the CMP-identity-mined hard-negative **image** partner and its id
- `hard_c`: the hard-negative partner's caption
- `source_id`, `source_caption`: provenance back to the synthetic source

`test/attr.json` and `test/ucc.json`, `multi-weather/*.json`, and `source_caption.json` are the
corresponding files for the test/multi-weather splits and the shared caption-source index.

This is **only the annotation metadata**, not image pixels. The PAB imagery itself is obtained
separately (see the main repo's README, "Data setup") from the AI City Challenge Track 4 organizers
or the dataset authors' own release; it is not redistributed here.

## Usage

```python
from huggingface_hub import snapshot_download
annotation_dir = snapshot_download(repo_id="Abdrah/scout-eccv-pab-annotations", repo_type="dataset")
```

Then point a SCOUT training config or `scripts/local_eval.py --hard-neg-annotation-dir` at
`{annotation_dir}/train` (or `/test`).

## License and provenance

The annotation content originates from the PAB dataset authors' CMP release
([Shuyu-XJTU/CMP](https://github.com/Shuyu-XJTU/CMP), MIT License, Copyright (c) 2025 Shuyu-XJTU)
and is redistributed here under the same MIT License, unmodified except for restoring the fields
the Track 4 release had stripped. All credit for the dataset and its annotations goes to the PAB
authors.

## Citation

If you use these annotations, cite the PAB/CMP paper:

```bibtex
@inproceedings{yang2025beyondwalking,
  title     = {Beyond walking: A large-scale image-text benchmark for text-based person anomaly search},
  author    = {Yang, Shuyu and Wang, Yaxiong and Zhu, Li and Zheng, Zhedong},
  booktitle = {ICCV},
  year      = {2025},
  note      = {arXiv:2411.17776}
}
```