File size: 4,439 Bytes
83b29f0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
---
pretty_name: TinyVIRAT (original 2020 release)
license: unknown
task_categories:
- video-classification
tags:
- tinyvirat
- virat
- action-recognition
- multi-label
- surveillance
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
  - split: test
    path: data/test-*.parquet
---

# TinyVIRAT

Hugging Face packaging of the **original TinyVIRAT release**, not TinyVIRAT-v2.
Natural low-resolution action clips extracted from VIRAT surveillance footage.
Each clip can have multiple action labels.

- [Original archive](https://www.crcv.ucf.edu/datasets/ugur/TinyVIRAT.zip)
- [Authors' repository](https://github.com/UgurDemir/Tiny-VIRAT)
- [Paper](https://arxiv.org/abs/2007.07355)

## Splits

| Split | Clips |
|---|---:|
| train | 7,663 |
| test | 5,166 |

These are the archive's original splits and annotation order. No validation
split is invented. IDs are local to each split. Group by `video_id` when
creating validation folds to avoid putting related surveillance clips in both
training and validation data.

## Usage

```python
from datasets import load_dataset

# Replace this local directory with the Hub repository ID after uploading.
ds = load_dataset("./hf_dataset")
example = ds["train"][0]
print(example["label_names"])
video_bytes = example["video"]["bytes"]
class_names = ds["train"].features["labels"].feature.names
```

The `video` feature defaults to `decode=False`, so metadata and raw bytes work
without PyTorch or TorchCodec. To decode, install compatible PyTorch,
TorchCodec and FFmpeg versions, then:

```python
from datasets import Video

decoded = ds["train"].cast_column("video", Video())
frame = decoded[0]["video"].get_frame_at(0).data
```

Streaming is supported: `load_dataset("./hf_dataset", streaming=True)`.

## Fields

| Field | Meaning |
|---|---|
| `id` | Original integer tube ID (unique within its split). |
| `video_id` | Original source surveillance-video identifier. |
| `path` | Original relative MP4 path from the annotation. |
| `video` | HF `Video` feature containing the original MP4 bytes and a relative path hint. |
| `labels` | List of zero-based `ClassLabel` IDs; this is a multi-label task. |
| `label_names` | Original action strings, preserving spelling, case and order. |
| `dim` | Original three annotation dimensions; first is frame count, followed by spatial dimensions. |

All clips in this release are square. Videos are neither resized nor re-encoded.
MP4 bytes are embedded in each Parquet shard: no dependency on an extracted
archive, local absolute paths, or upstream URLs at load time. Original
annotation JSON is retained in `annotations/`.

## Classes

IDs follow lexicographic ordering of the 26 original strings. This wrapper's
numeric mapping is not an upstream-provided class ordering. It is also stored
in `labels.json` and in the Parquet feature metadata.

| ID | Label |
|---:|---|
| 0 | `Closing` |
| 1 | `Entering` |
| 2 | `Exiting` |
| 3 | `Interacts` |
| 4 | `Loading` |
| 5 | `Misc` |
| 6 | `Opening` |
| 7 | `Pull` |
| 8 | `Push` |
| 9 | `Riding` |
| 10 | `Talking` |
| 11 | `Transport_HeavyCarry` |
| 12 | `activity_carrying` |
| 13 | `activity_gesturing` |
| 14 | `activity_running` |
| 15 | `activity_standing` |
| 16 | `activity_walking` |
| 17 | `specialized_miscellaneous` |
| 18 | `specialized_talking_phone` |
| 19 | `specialized_texting_phone` |
| 20 | `specialized_using_tool` |
| 21 | `vehicle_moving` |
| 22 | `vehicle_starting` |
| 23 | `vehicle_stopping` |
| 24 | `vehicle_turning_left` |
| 25 | `vehicle_turning_right` |

## Provenance and rights

Source ZIP SHA-256: `a04e34f3d74ead4103c696c4d8fb4b5dd856d15e6b748ba0b2864f3a2a8c520c`.
`provenance.json` records the input checksum and split sizes.

No explicit dataset redistribution license was found in the supplied archive
or the authors' repository. **Unknown is not permission to redistribute.**
Check the TinyVIRAT/VIRAT terms and obtain any necessary permissions before
publishing a mirror or using the footage. This wrapper does not grant new
rights to the source data. Surveillance footage can depict people and
vehicles; consider privacy and potential surveillance harms.

## Citation

```bibtex
@misc{demir2020tinyvirat,
  title={TinyVIRAT: Low-resolution Video Action Recognition},
  author={Ugur Demir and Yogesh S Rawat and Mubarak Shah},
  year={2020},
  eprint={2007.07355},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}
```