FelixKAI commited on
Commit
8b77017
·
0 Parent(s):

Duplicate from RL-MIND/NJU-HARD-Tracking

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +3 -0
  2. README.md +205 -0
  3. README_ZH.md +174 -0
  4. assets/dataset-logo-v2.webp +3 -0
  5. assets/dataset-logo.png +3 -0
  6. assets/rl-mind-logo-v2.webp +3 -0
  7. assets/rl-mind-logo.png +3 -0
  8. data/test-00000.parquet +3 -0
  9. data/test-00001.parquet +3 -0
  10. data/test-00002.parquet +3 -0
  11. data/test-00003.parquet +3 -0
  12. data/test-00004.parquet +3 -0
  13. data/test-00005.parquet +3 -0
  14. data/test-00006.parquet +3 -0
  15. data/test-00007.parquet +3 -0
  16. data/test-00008.parquet +3 -0
  17. data/test-00009.parquet +3 -0
  18. data/test-00010.parquet +3 -0
  19. data/test-00011.parquet +3 -0
  20. data/test-00012.parquet +3 -0
  21. data/test-00013.parquet +3 -0
  22. data/test-00014.parquet +3 -0
  23. data/test-00015.parquet +3 -0
  24. data/test-00016.parquet +3 -0
  25. data/test-00017.parquet +3 -0
  26. data/test-00018.parquet +3 -0
  27. data/test-00019.parquet +3 -0
  28. data/test-00020.parquet +3 -0
  29. data/test-00021.parquet +3 -0
  30. data/test-00022.parquet +3 -0
  31. data/test-00023.parquet +3 -0
  32. data/test-00024.parquet +3 -0
  33. data/test-00025.parquet +3 -0
  34. data/test-00026.parquet +3 -0
  35. data/test-00027.parquet +3 -0
  36. data/test-00028.parquet +3 -0
  37. data/test-00029.parquet +3 -0
  38. data/test-00030.parquet +3 -0
  39. data/test-00031.parquet +3 -0
  40. data/test-00032.parquet +3 -0
  41. data/test-00033.parquet +3 -0
  42. data/test-00034.parquet +3 -0
  43. data/test-00035.parquet +3 -0
  44. data/test-00036.parquet +3 -0
  45. data/test-00037.parquet +3 -0
  46. data/test-00038.parquet +3 -0
  47. data/test-00039.parquet +3 -0
  48. data/test-00040.parquet +3 -0
  49. data/test-00041.parquet +3 -0
  50. data/test-00042.parquet +3 -0
.gitattributes ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ *.parquet filter=lfs diff=lfs merge=lfs -text
2
+ *.png filter=lfs diff=lfs merge=lfs -text
3
+ *.webp filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,205 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: unknown
3
+ pretty_name: NJU-HARD-Tracking
4
+ task_categories:
5
+ - other
6
+ tags:
7
+ - remote-sensing
8
+ - uav
9
+ - aerial-imagery
10
+ - ultra-high-resolution
11
+ - parquet
12
+ - multi-object-tracking
13
+ - object-tracking
14
+ size_categories:
15
+ - 1K<n<10K
16
+ configs:
17
+ - config_name: default
18
+ default: true
19
+ data_files:
20
+ - split: train
21
+ path: data/train-*.parquet
22
+ - split: validation
23
+ path: data/validation-*.parquet
24
+ - split: test
25
+ path: data/test-*.parquet
26
+ - config_name: preview
27
+ data_files:
28
+ - split: sample
29
+ path: preview/*-preview.parquet
30
+ ---
31
+
32
+ # NJU-HARD-Tracking
33
+
34
+ <table align="center" role="presentation" style="margin:0 auto; border:0; background:transparent;">
35
+ <tr style="border:0; background:transparent;">
36
+ <td align="center" style="border:0; background:transparent; padding:0 8px;">
37
+ <img src="./assets/rl-mind-logo-v2.webp" alt="RL-MIND research group logo" width="160" height="160" loading="eager" fetchpriority="high" decoding="async" style="width:160px; height:160px; object-fit:contain;" />
38
+ </td>
39
+ <td align="center" style="border:0; background:transparent; padding:0 8px;">
40
+ <img src="./assets/dataset-logo-v2.webp" alt="NJU-HARD-Tracking dataset logo" width="160" height="160" loading="eager" fetchpriority="high" decoding="async" style="width:160px; height:160px; object-fit:contain;" />
41
+ </td>
42
+ </tr>
43
+ </table>
44
+
45
+ <p align="center"><b>Multi-Object Tracking across 122 MP UAV Image Sequences</b></p>
46
+
47
+ <p align="center"><a href="https://huggingface.co/datasets/RL-MIND/NJU-HARD-Tracking">🤗 Hugging Face</a> · <a href="https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD-Tracking">🟣 ModelScope</a> · 📊 Statistics: <a href="https://huggingface.co/datasets/RL-MIND/NJU-HARD-Tracking/blob/main/statistics.json">HF</a> / <a href="https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD-Tracking/file/view/master/statistics.json?status=1">MS</a></p>
48
+
49
+ <p align="center"><b>English</b> | 中文: <a href="https://huggingface.co/datasets/RL-MIND/NJU-HARD-Tracking/blob/main/README_ZH.md">Hugging Face</a> · <a href="https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD-Tracking/file/view/master/README_ZH.md?status=1">ModelScope</a></p>
50
+
51
+ ## 🌍 Overview
52
+
53
+ **NJU-HARD-Tracking** provides the **multi-object-tracking release of HARD**, with full-resolution frames, temporal ordering, and the original instance annotations. It supports studying how detection and association behave across wide-area aerial scenes.
54
+
55
+ The source contains **3,549 original JPEG frames**, **21 sequences**, and **138,830 box annotations** across four classes: pedestrian, car, bus, and truck. Every frame is **12,768 × 9,564 pixels** (approximately **122 MP**). The imagery covers sparse and dense scenes, with substantial variation in target scale and camera motion.
56
+
57
+ HARD stands for **Ultra-High-resolution Airborne Remote-sensing Dataset**. The [NJU-HARD-Detection release](https://huggingface.co/datasets/RL-MIND/NJU-HARD-Detection) contains the **same frame and annotation records** with a task-specific card. Both releases preserve the original JPEG bytes and labels. Their repository names identify these task releases; they do not denote separate acquisitions.
58
+
59
+ ## ✨ What the Data Covers
60
+
61
+ - **Wide-area detail:** complete 122 MP frames retain small pedestrian and vehicle targets within a broad field of view.
62
+ - **Temporal structure:** 21 ordered sequences connect per-frame localization with cross-frame association.
63
+ - **Diverse viewing conditions:** the source manuscript describes 50–355 m acquisition heights and four flight modes: linear-nadir, linear-oblique, stationary-gaze, and orbit-gaze.
64
+ - **Realistic imbalance:** cars dominate the annotations, while buses and trucks are less frequent. Scene density ranges from approximately 12.4 to 79.8 annotated objects per frame across the three capture subsets.
65
+
66
+ These properties support research on wide-area spatio-temporal scene understanding. They do not establish a performance claim for any model.
67
+
68
+ ## 📊 Dataset at a Glance
69
+
70
+ | Item | Source inventory represented by the release |
71
+ |:---|:---|
72
+ | Original frames | 3,549 |
73
+ | Image resolution | 12,768 × 9,564 pixels |
74
+ | Capture subsets / sequences | 3 / 21 |
75
+ | Box annotation rows | 138,830 |
76
+ | Frames with a source label file | 3,544 |
77
+ | Frames without a source label file | 5; annotation state is missing |
78
+ | Source JPEG payload | 368,182,969,599 bytes, approximately 368.18 GB |
79
+ | Storage unit | One original frame per Parquet row |
80
+ | Configuration | `default` |
81
+ | Splits | `train`, `validation`, `test` |
82
+ | Separate browsing preview | `preview` / `sample`: 7 thumbnail frames in separate Parquet files, at most 1,024 pixels on the longest edge |
83
+
84
+ The JPEG byte count describes the original image payload, not the final Parquet file size. In `default`, each image embeds its original `bytes` and has `path=null`; it is not resized, cropped, or re-encoded. The separate seven-frame `preview` uses derived thumbnails and is not included in the 3,549-frame total. It spans the three subsets and splits, including examples with missing labels, repeated identities, and out-of-bounds boxes. Box totals count annotation rows across frames, not distinct physical objects or verified unique trajectories.
85
+
86
+ | Dense `category` | Class | Original `source_class_id` | Annotation rows |
87
+ |---:|:---|---:|---:|
88
+ | 0 | pedestrian | 0 | 27,711 |
89
+ | 1 | car | 2 | 103,361 |
90
+ | 2 | bus | 5 | 4,033 |
91
+ | 3 | truck | 7 | 3,725 |
92
+
93
+ The four source-ID totals match the manuscript's per-class totals exactly. `category` provides a contiguous label space; `source_class_id` preserves the original 0/2/5/7 values.
94
+
95
+ ## 🗂️ Sequence Splits
96
+
97
+ The releases use the sequence-level split in the manuscript's Table 4. All frames from a sequence stay together. The mapping from the manuscript's A/B/C subset codes to source directories was reconstructed from sequence inventories and verified by exact agreement with all three split annotation totals.
98
+
99
+ | Split | Sequences | Frames | Annotation rows |
100
+ |:---|---:|---:|---:|
101
+ | `train` | 13 | 2,055 | 89,148 |
102
+ | `validation` | 3 | 551 | 15,820 |
103
+ | `test` | 5 | 943 | 33,862 |
104
+
105
+ | Source directory / paper code | `train` sequence IDs | `validation` | `test` |
106
+ |:---|:---|:---|:---|
107
+ | `UAV-BASE-13` / A | 1, 2, 3, 4, 5 | 6 | — |
108
+ | `UAV-BASE-14` / B | 1, 2, 5, 8, 9 | 3, 7 | 4, 6 |
109
+ | `NJU-xianlin` / C | 1, 5, 6 | — | 2, 3, 4 |
110
+
111
+ The split is **sequence-disjoint, not site-disjoint**. The manuscript's tracking evaluation pools the **five test sequences** with TrackEval's combined convention. The complete tracking release includes all 21 sequences for training, validation, and testing.
112
+
113
+ ## 🧩 Record Format
114
+
115
+ Each Parquet row is one frame. The detection and tracking releases share this schema so that image bytes, labels, and source provenance remain comparable.
116
+
117
+ | Field | Meaning |
118
+ |:---|:---|
119
+ | `image`, `preview_only` | Native image value with a null path; original JPEG in `default`, a derived thumbnail when `preview_only=true` |
120
+ | `image_id`, `scene`, `sequence_id` | Frame identity, capture subset, and sequence identity |
121
+ | `source_sequence_id`, `paper_sequence_id` | Original sequence number and corresponding A/B/C manuscript sequence code |
122
+ | `frame_id`, `frame_name` | Integer parsed from the original frame filename, and that filename; the original starting index is retained |
123
+ | `split`, `width`, `height` | Assigned source-sequence split and dimensions of the current `image.bytes` |
124
+ | `source_width`, `source_height` | Original full-resolution dimensions, also retained in the preview |
125
+ | `source_image_path`, `source_image_sha256` | Relative source path and original JPEG checksum |
126
+ | `image_sha256` | Checksum of the current `image.bytes`; identical to the source checksum in `default`, different for a derived preview |
127
+ | `annotation_status`, `num_objects` | `present` or `missing`; object count is null when the source label is missing |
128
+ | `source_annotation_path`, `raw_annotation_bytes`, `annotation_sha256` | Original label provenance and exact bytes; bytes/checksum are null for missing labels |
129
+ | `quality_flags` | Frame-level source-quality indicators |
130
+ | `objects` | List of object records, or **null** for a missing source label |
131
+
132
+ Each object includes:
133
+
134
+ | Field | Meaning |
135
+ |:---|:---|
136
+ | `id`, `source_line` | Source label path plus row identity, and the original line number |
137
+ | `category`, `source_class_id` | Contiguous class label and unchanged source class ID |
138
+ | `source_track_id`, `source_track_key` | Raw instance ID and its sequence/class-scoped key |
139
+ | `source_bbox_normalized` | Original YOLO-style `[cx, cy, width, height]` in normalized coordinates |
140
+ | `bbox`, `area` | COCO-style `[x_min, y_min, width, height]` and area in the current image's pixels, retained as floating point |
141
+ | `quality_flags` | Object-level source-quality indicators |
142
+
143
+ For tracking, process frames within each `sequence_id` in increasing `frame_id` order. `source_track_key` includes the sequence, source class, and raw instance ID; it prevents cross-class ID collisions but does not resolve the same-class ambiguities listed below. Preserve `source_track_id` when exporting or comparing source labels.
144
+
145
+ Pixel boxes are derived from the original normalized center/size coordinates and the current `width`/`height`. They use original-image pixels in `default` and thumbnail pixels in `preview`; they are **not clipped or rounded to integers**. No actual capture timestamp or measured video frame rate is inferred from the frame index. The manuscript's `P = 2 s` is recorded as evaluation context, not synthesized per-frame timing metadata.
146
+
147
+ ## ⚠️ Source Annotation Quality
148
+
149
+ The release preserves the audited source rather than silently changing ambiguous annotations. Quality flags and original label bytes make the following issues visible:
150
+
151
+ | Source issue | Count / interpretation |
152
+ |:---|:---|
153
+ | Missing label files | 5 frames; `objects=null`, not an empty list or a confirmed negative frame |
154
+ | Raw IDs reused across classes within a sequence | 158 sequence/ID groups; do not use raw ID alone as a track key |
155
+ | Same class and ID repeated within a frame | 24 groups; a class-scoped key still has source ambiguity in these cases |
156
+ | Exact duplicate annotation row | 1 duplicate occurrence, retained |
157
+ | Box corners beyond normalized image bounds | 2,149 rows at a tolerance of `1e-6`; original coordinates retained |
158
+
159
+ The missing label files correspond to `UAV-BASE-13` sequence 1 frame 0, sequence 2 frame 0, and sequence 6 frames 0–2. There are **no empty source TXT files**. Downstream experiments should state how they treat missing labels, duplicate identities, duplicate rows, and out-of-bounds boxes. A cleaned evaluation protocol must document its transformations separately; the source track keys are not a claim of fully resolved ground-truth identities.
160
+
161
+ ## 🚀 Quick Start
162
+
163
+ Start with the small `preview` to inspect the schema and quality flags. With Hugging Face `datasets`, streaming and disabled image decoding keep this inspection lightweight:
164
+
165
+ ```python
166
+ from datasets import Image, load_dataset
167
+
168
+ data = load_dataset(
169
+ "RL-MIND/NJU-HARD-Tracking",
170
+ name="preview",
171
+ split="sample",
172
+ streaming=True,
173
+ ).cast_column("image", Image(decode=False))
174
+
175
+ frame = next(iter(data))
176
+ print(frame["sequence_id"], frame["frame_id"], frame["annotation_status"])
177
+ if frame["objects"] is None:
178
+ print("Missing source label; this is not a negative frame.")
179
+ else:
180
+ for obj in frame["objects"]:
181
+ print(obj["source_track_key"], obj["bbox"], obj["quality_flags"])
182
+ ```
183
+
184
+ ModelScope's dataset loader exposes the same preview:
185
+
186
+ ```python
187
+ from modelscope.msdatasets import MsDataset
188
+
189
+ data = MsDataset.load(
190
+ "KAIWANG/NJU-HARD-Tracking",
191
+ subset_name="preview",
192
+ split="sample",
193
+ use_streaming=True,
194
+ )
195
+ frame = next(iter(data))
196
+ print(frame["sequence_id"], frame["frame_id"], frame["annotation_status"])
197
+ ```
198
+
199
+ For original frames, use `name="default", split="test"` on Hugging Face, or `subset_name="default", split="test"` on ModelScope; `train` and `validation` are also available. The preview is a browsing aid, **not a continuous tracking evaluation sequence**. Use `default` for full-resolution experiments and chronological tracking. Full-resolution frames remain large even when streamed; after resizing an image, transform its boxes consistently.
200
+
201
+ ## 📄 License and Research Context
202
+
203
+ License metadata is **unknown**. The source manuscript describes research-only distribution and prohibits attempts to re-identify individuals or vehicles, but a complete license agreement is not supplied with the source material reviewed for this release.
204
+
205
+ The detection task uses per-frame box supervision; the tracking task additionally studies sequence-local association. The manuscript reports COCO detection metrics and HOTA/s-HOTA tracking evaluation. This data package does not, by itself, establish a validated implementation of the manuscript's latency-accounting protocol.
README_ZH.md ADDED
@@ -0,0 +1,174 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # NJU-HARD-Tracking
2
+
3
+ <table align="center" role="presentation" style="margin:0 auto; border:0; background:transparent;">
4
+ <tr style="border:0; background:transparent;">
5
+ <td align="center" style="border:0; background:transparent; padding:0 8px;">
6
+ <img src="./assets/rl-mind-logo-v2.webp" alt="RL-MIND research group logo" width="160" height="160" loading="eager" fetchpriority="high" decoding="async" style="width:160px; height:160px; object-fit:contain;" />
7
+ </td>
8
+ <td align="center" style="border:0; background:transparent; padding:0 8px;">
9
+ <img src="./assets/dataset-logo-v2.webp" alt="NJU-HARD-Tracking dataset logo" width="160" height="160" loading="eager" fetchpriority="high" decoding="async" style="width:160px; height:160px; object-fit:contain;" />
10
+ </td>
11
+ </tr>
12
+ </table>
13
+
14
+ <p align="center"><b>面向 122 MP 无人机影像序列的多目标跟踪</b></p>
15
+
16
+ <p align="center"><a href="https://huggingface.co/datasets/RL-MIND/NJU-HARD-Tracking">🤗 Hugging Face</a> · <a href="https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD-Tracking">🟣 ModelScope</a> · 📊 Statistics: <a href="https://huggingface.co/datasets/RL-MIND/NJU-HARD-Tracking/blob/main/statistics.json">HF</a> / <a href="https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD-Tracking/file/view/master/statistics.json?status=1">MS</a></p>
17
+
18
+ <p align="center"><b>中文</b> | English: <a href="https://huggingface.co/datasets/RL-MIND/NJU-HARD-Tracking">Hugging Face</a> · <a href="https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD-Tracking">ModelScope</a></p>
19
+
20
+ ## 🌍 数据集简介
21
+
22
+ **NJU-HARD-Tracking** 是 HARD 的**多目标跟踪发布版本**,提供原始分辨率影像、时间顺序及原始实例标注,支持研究广域航拍场景中的目标检测与跨帧关联。
23
+
24
+ 源数据包含 **3,549 张原始 JPEG 影像**、**21 个序列**和 **138,830 行目标框标注**,覆盖行人、小汽车、公交车和卡车四类目标。所有影像均为 **12,768 × 9,564 像素**,约 **122 MP**。不同采集子集覆盖稀疏和密集场景,具有明显的目标尺度及相机运动差异。
25
+
26
+ HARD 的全称为 **Ultra-High-resolution Airborne Remote-sensing Dataset**。[NJU-HARD-Detection 版本](https://huggingface.co/datasets/RL-MIND/NJU-HARD-Detection) 包含**相同的影像及标注记录**,并提供对应任务的说明。两个版本均保留原始 JPEG 字节和标签;仓库名称区分任务发布形式,不表示两次独立采集。
27
+
28
+ ## ✨ 数据特色
29
+
30
+ - **广域细节**:完整的 122 MP 影像在覆盖广阔视野的同时保留小尺度行人与车辆目标。
31
+ - **时序结构**:21 个有序序列将逐帧定位与跨帧身份关联连接起来。
32
+ - **多种观测条件**:源论文描述了 50–355 m 的采集高度,以及线性俯视、线性斜视、定点凝视和环绕凝视四种飞行模式。
33
+ - **自然类别不均衡**:小汽车占比较高,公交车和卡车相对稀少。三个采集子集的平均目标密度约为每帧 12.4 至 79.8 个目标。
34
+
35
+ 这些特征服务于广域时空场景理解研究,不构成任何模型性能结论。
36
+
37
+ ## 📊 数据概览
38
+
39
+ | 项目 | 本版本表示的源数据清单 |
40
+ |:---|:---|
41
+ | 原始影像帧数 | 3,549 |
42
+ | 图像分辨率 | 12,768 × 9,564 像素 |
43
+ | 采集子集 / 序列数 | 3 / 21 |
44
+ | 目标框标注行数 | 138,830 |
45
+ | 有源标签文件的帧 | 3,544 |
46
+ | 缺少源标签文件的帧 | 5;标注状态为缺失 |
47
+ | 源 JPEG 数据量 | 368,182,969,599 字节,约 368.18 GB |
48
+ | 存储粒度 | 每个 Parquet 行对应一张原始帧 |
49
+ | 配置 | `default` |
50
+ | 数据划分 | `train`、`validation`、`test` |
51
+ | 独立浏览预览 | `preview` / `sample`:7 张缩略图,分别存于独立 Parquet 文件,最长边不超过 1,024 像素 |
52
+
53
+ JPEG 字节数表示原始图像数据量,不是最终 Parquet 文件大小。`default` 每行图像直接内嵌原始 `bytes`,其 `path=null`;原图不缩放、不裁剪、不重新编码。独立的七帧 `preview` 使用派生缩略图,不计入 3,549 张原图总数;预览覆盖三个采集子集和三个 split,并包含缺标签、身份重复及越界框示例。目标框总数按跨帧标注行计数,不是去重后的现实目标数或已核实的唯一轨迹数。
54
+
55
+ | 连续 `category` | 类别 | 原始 `source_class_id` | 标注行数 |
56
+ |---:|:---|---:|---:|
57
+ | 0 | pedestrian(行人) | 0 | 27,711 |
58
+ | 1 | car(小汽车) | 2 | 103,361 |
59
+ | 2 | bus(公交车) | 5 | 4,033 |
60
+ | 3 | truck(卡车) | 7 | 3,725 |
61
+
62
+ 四个源类别 ID 的全量计数与论文对应类别的统计精确匹配。`category` 提供连续类别标签,`source_class_id` 保留原始的 0/2/5/7 编号。
63
+
64
+ ## 🗂️ 序列划分
65
+
66
+ 两个版本采用论文 Table 4 的序列级划分,同一序列的全部帧保留在同一 split。论文 A/B/C 子集代码与源目录之间的对应关系由序列清单重建,并通过三个 split 的目标框总数全部精确吻合得到验证。
67
+
68
+ | Split | 序列数 | 帧数 | 标注行数 |
69
+ |:---|---:|---:|---:|
70
+ | `train` | 13 | 2,055 | 89,148 |
71
+ | `validation` | 3 | 551 | 15,820 |
72
+ | `test` | 5 | 943 | 33,862 |
73
+
74
+ | 源目录 / 论文代码 | `train` 序列 ID | `validation` | `test` |
75
+ |:---|:---|:---|:---|
76
+ | `UAV-BASE-13` / A | 1, 2, 3, 4, 5 | 6 | — |
77
+ | `UAV-BASE-14` / B | 1, 2, 5, 8, 9 | 3, 7 | 4, 6 |
78
+ | `NJU-xianlin` / C | 1, 5, 6 | — | 2, 3, 4 |
79
+
80
+ 该划分**序列互斥,但不是地点互斥**。论文的跟踪评价通过 TrackEval 的 combined 规则聚合**五个测试序列**。完整跟踪发布版本包含全部 21 个序列,分别用于训练、验证和测试。
81
+
82
+ ## 🧩 记录格式
83
+
84
+ 每个 Parquet 行对应一张影像帧。检测和跟踪版本共用这一 schema,以便核对原图字节、标签和来源信息。
85
+
86
+ | 字段 | 含义 |
87
+ |:---|:---|
88
+ | `image`、`preview_only` | 路径为空的原生图像字段;`default` 是原始 JPEG,`preview_only=true` 时为派生缩略图 |
89
+ | `image_id`、`scene`、`sequence_id` | 帧身份、采集子集及序列身份 |
90
+ | `source_sequence_id`、`paper_sequence_id` | 原始序列编号及对应的论文 A/B/C 序列代码 |
91
+ | `frame_id`、`frame_name` | 从原始文件名解析的整数帧编号及该文件名;保留源序列的起始索引 |
92
+ | `split`、`width`、`height` | 源序列所属 split 与当前 `image.bytes` 的图像尺寸 |
93
+ | `source_width`、`source_height` | 原始完整分辨率,预览中也保留 |
94
+ | `source_image_path`、`source_image_sha256` | 相对源路径及原始 JPEG 校验和 |
95
+ | `image_sha256` | 当前 `image.bytes` 的校验和;`default` 与原图相同,派生预览与原图不同 |
96
+ | `annotation_status`、`num_objects` | `present` 或 `missing`;缺少源标签时目标数为空 |
97
+ | `source_annotation_path`、`raw_annotation_bytes`、`annotation_sha256` | 源标签溯源及完整字节;缺标签时字节和校验和为空 |
98
+ | `quality_flags` | 帧级源数据质量标记 |
99
+ | `objects` | 目标记录列表;缺少源标签时为 **null** |
100
+
101
+ 每个目标包含:
102
+
103
+ | 字段 | 含义 |
104
+ |:---|:---|
105
+ | `id`、`source_line` | 源标签路径与行身份,以及原始行号 |
106
+ | `category`、`source_class_id` | 连续类别标签及未改写的源类别编号 |
107
+ | `source_track_id`、`source_track_key` | 原始实例 ID 及限定到序列和类别的键 |
108
+ | `source_bbox_normalized` | 原始 YOLO 格式归一化 `[cx, cy, width, height]` |
109
+ | `bbox`、`area` | 当前图像像素坐标下的 COCO 格式 `[x_min, y_min, width, height]` 及面积,保留浮点值 |
110
+ | `quality_flags` | 目标级源数据质量标记 |
111
+
112
+ 目标跟踪应在每个 `sequence_id` 内按 `frame_id` 递增顺序处理。`source_track_key` 同时包含序列、源类别和原始实例 ID,可避免跨类别 ID 冲突,但不能自动消除下文列出的同类别身份歧义。导出或核对源标签时应保留 `source_track_id`。
113
+
114
+ 像素框由原始归一化中心坐标和宽高,以及当前 `width`/`height` 转换:`default` 使用原图像素,`preview` 使用缩略图像素,**不裁剪、不取整为整数**。不会根据帧编号推算真实采集时间戳或实测视频帧率。论文的 `P = 2 s` 仅作为评价背景,不被写成推定的逐帧时间元数据。
115
+
116
+ ## ⚠️ 源标注质量
117
+
118
+ 本版本保留经过清点的源数据,不静默修改存在歧义的标注。质量标记和原始标签字节用于公开以下情况:
119
+
120
+ | 源数据问题 | 数量 / 解释 |
121
+ |:---|:---|
122
+ | 缺少标签文件 | 5 帧;`objects=null`,不是空列表或确认无目标的负样本 |
123
+ | 同一序列的原始 ID 跨类别复用 | 158 组序列/ID;不能只用原始 ID 作为轨迹键 |
124
+ | 同帧同类别同 ID 重复 | 24 组;即使用类别限定的键,这些标注仍有源身份歧义 |
125
+ | 完全重复的标注行 | 1 次重复,原样保留 |
126
+ | 框角点超出归一化图像边界 | 在 `1e-6` 容差下有 2,149 行;保留原始坐标 |
127
+
128
+ 缺标签帧为 `UAV-BASE-13` 的序列 1 第 0 帧、序列 2 第 0 帧,以及序列 6 第 0–2 帧。源数据中**没有空 TXT 文件**。下游实验应说明对缺标签、重复身份、重复行和越界框的处理方式。若建立清洗后的评价协议,应单独记录所有变换;源轨迹键不代表身份歧义已经全部消除。
129
+
130
+ ## 🚀 快速使用
131
+
132
+ 可以先使用体积较小的 `preview` 查看 schema 和质量标记。通过 Hugging Face `datasets` 流式读取并关闭图像解码,可减少检查时的开销:
133
+
134
+ ```python
135
+ from datasets import Image, load_dataset
136
+
137
+ data = load_dataset(
138
+ "RL-MIND/NJU-HARD-Tracking",
139
+ name="preview",
140
+ split="sample",
141
+ streaming=True,
142
+ ).cast_column("image", Image(decode=False))
143
+
144
+ frame = next(iter(data))
145
+ print(frame["sequence_id"], frame["frame_id"], frame["annotation_status"])
146
+ if frame["objects"] is None:
147
+ print("Missing source label; this is not a negative frame.")
148
+ else:
149
+ for obj in frame["objects"]:
150
+ print(obj["source_track_key"], obj["bbox"], obj["quality_flags"])
151
+ ```
152
+
153
+ ModelScope 数据集加载器提供同样的预览���
154
+
155
+ ```python
156
+ from modelscope.msdatasets import MsDataset
157
+
158
+ data = MsDataset.load(
159
+ "KAIWANG/NJU-HARD-Tracking",
160
+ subset_name="preview",
161
+ split="sample",
162
+ use_streaming=True,
163
+ )
164
+ frame = next(iter(data))
165
+ print(frame["sequence_id"], frame["frame_id"], frame["annotation_status"])
166
+ ```
167
+
168
+ 读取原图时,Hugging Face 使用 `name="default", split="test"`,ModelScope 使用 `subset_name="default", split="test"`;也可选择 `train` 或 `validation`。预览用于浏览检查,**不是连续的跟踪评价序列**;完整分辨率实验和按时间顺序的跟踪应使用 `default`。即使采用流式读取,单张原始分辨率影像仍较大;缩放影像时,需要同步变换对应目标框。
169
+
170
+ ## 📄 许可与研究背景
171
+
172
+ 许可元数据记为 **unknown**。源论文描述了仅限研究用途的分发方式,并禁止尝试重新识别个人或车辆;但本次审阅的源材料中未提供完整许可协议。
173
+
174
+ 检测任务使用逐帧框监督,跟踪任务进一步研究序列内身份关联。论文报告了 COCO 检测指标以及 HOTA/s-HOTA 跟踪评价。此数据包本身不等于已经验证了论文的延迟统计协议实现。
assets/dataset-logo-v2.webp ADDED

Git LFS Details

  • SHA256: d671335a5a553cebd64b12d3896e516781704d335944af02636b44c23909994b
  • Pointer size: 130 Bytes
  • Size of remote file: 50.5 kB
assets/dataset-logo.png ADDED

Git LFS Details

  • SHA256: 8adbe08f5a2e4ee0c414236d05df40aea497624491771feea4c94d6b080e70d5
  • Pointer size: 132 Bytes
  • Size of remote file: 2.1 MB
assets/rl-mind-logo-v2.webp ADDED

Git LFS Details

  • SHA256: 56d2d9f54d2359041cb3c86cab7fe825b0bf9e3a3eed672b815bcd91669d3a46
  • Pointer size: 129 Bytes
  • Size of remote file: 4.66 kB
assets/rl-mind-logo.png ADDED

Git LFS Details

  • SHA256: 89531f0412eae024bd022e68c197ba6fb822a6fc036589393dd1ca39679916ba
  • Pointer size: 130 Bytes
  • Size of remote file: 12.9 kB
data/test-00000.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:459320dcd5d26ddfe9a9ceea0c2457597419b75359ab107e6cc76dca802ee468
3
+ size 795266290
data/test-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ec5c7e617e3eb6004defd6ed3df05ec1bf1efae6a3e8fc415c4a9cf5738307b1
3
+ size 787493075
data/test-00002.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:77e13375e578553fc357acb7f900fa3c34fccdd750ab4a58b110a438958e725d
3
+ size 777536610
data/test-00003.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:73068557244cc62e327ca06c436b4ec1a4dd88944a1c923dd4cf9267fe8fdf20
3
+ size 769287403
data/test-00004.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:686146cb1f65d3d9b52b0bec9525bd594dd3ef7f2b585fb48bc741059dab13ff
3
+ size 762349763
data/test-00005.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dc721e0e9e38f80de30e79d82d5d7009340ba0912572d2169d0845dd578d2f93
3
+ size 761528506
data/test-00006.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:311e9cfaffe6c3658b4d352f64247291830a8005dbbe37e05f005cd3389944e5
3
+ size 758714159
data/test-00007.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8e401c36273dd9abb3c9192d2bf37b5d4a0bbeffec25b90dd1b14b7ea1828af5
3
+ size 761563787
data/test-00008.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:209cc3701b5e640403cebc1ea1b9cde14c0ef763be8bb6d11af71782f004fbbe
3
+ size 775555887
data/test-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8c10ee8db9f31edf5af0e5626dee03a34cc054092c53483d6b5ab066d1c3aa8e
3
+ size 787532835
data/test-00010.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c24f9a999aad71c3ef9e4b13f80e77bf03064c7bc7cd18388b5cd300d2ce2c64
3
+ size 796593398
data/test-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:20e577195fb5df006ba87c0cc795ac41aa3e3ae1cdd0c9bb1b9666f30f8e46f7
3
+ size 706682310
data/test-00012.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b835b7d6a47224e2b2b6ade5db406e687843f870b0ecf3836721da12579161de
3
+ size 714437824
data/test-00013.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6bcc70c5e95f8e19630ceedcdfcd124414248cd7fae147a9859d359685a1fb7e
3
+ size 721624747
data/test-00014.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c20573b049420b2a99b4e8f22454618c1ce69ef4cf46efe5923ae3779b81fe8f
3
+ size 415486543
data/test-00015.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1817b48ec42ec690fa1437cb64da1b9c813e3de6e59e7e253ed2b2c1b0a54348
3
+ size 746946095
data/test-00016.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e094b93a7158049c04211e287558ba22c0a9dc1451f1d6d137755f164460f6d
3
+ size 743726601
data/test-00017.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b829b8f640a1fa1f1725569388d851d8e35f522e729637ea0ecaafd6393c3cb1
3
+ size 742359986
data/test-00018.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ce3f8bd3b5993d67ad80cdb545686c7e91e7edb5f98e6c6511f60b7245f1f84d
3
+ size 745210289
data/test-00019.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bcd92dc6c81bf5f526206ced40bde8cff5d9d29f2074c1d13187231e8d15af53
3
+ size 749136450
data/test-00020.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6fbec18c88c4511676919a324eab7e74213e8f9e684c1dc25f8c75cd6ac13090
3
+ size 745177314
data/test-00021.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:615da74d914b62e125d827debf5f696a01212400143b02628e216e1ae95aee84
3
+ size 741194235
data/test-00022.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6f23301d6760d9ea8a1d46b148462650f598c089543df6bd56d9365b1d503af1
3
+ size 738383433
data/test-00023.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2f4a17b36190a9c6cf18e06fda1be3c4805f753d1e5083565ffdad8ba4802e89
3
+ size 736715425
data/test-00024.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:df7e3385f93d675a8be6bb8cfba008111fea328b1f522b8fea2d0b09aba2a47d
3
+ size 737074339
data/test-00025.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6cc1bb39cde879270f3ad8b3d997c115da4a38b4e2a7ed33d0741ab58a5f50b9
3
+ size 734777187
data/test-00026.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2cbc4b05cba2dd584771c5f272378d60014d0f8755eec20c723502e19b53a434
3
+ size 729545129
data/test-00027.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:510ca3de6222076c43f197ac0b1234ded27604e193601f286dabc7d5abf83aa3
3
+ size 724344695
data/test-00028.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c19c3dc5a4749dea835475245a39cd2f2c0179676bb9ebd3444849eeac5fd042
3
+ size 724500815
data/test-00029.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9876086ef0456e00af1d970c878273e1c44672f93cd2d36c536a8d75b28bacda
3
+ size 726969922
data/test-00030.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b55d5adffbbb5c590a33319f9f4e1f9ca57898cc253180f3fdff200b28c34f48
3
+ size 726032330
data/test-00031.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0aabd516af7cbab06b30b410c4495affbfd9d1d9128a6a86a7c7a4ba71d72670
3
+ size 724458203
data/test-00032.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b94b0920a8222fc28bf56e08e4669827432fe113fba28b9389bb0fb582dcafb5
3
+ size 721837736
data/test-00033.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3745e72e2689c4b40d8924a0a66012c40c60993f0d85f10fb660a860fd4d2ed1
3
+ size 718829289
data/test-00034.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:554c35b71bda544d6b35828459d3ae9f6f69997745ea3c053b44067eca5b7737
3
+ size 713273303
data/test-00035.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3ab6334baf6d091971ebe6188b6a56d84caf391d58c11b0a57ca45451a357db
3
+ size 803592671
data/test-00036.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b4bceb36e182e47652ae4a8a473d3aebb36a1186927acf2e7d79ca685781af4e
3
+ size 793937358
data/test-00037.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9db756442903c6875cb97f64c89684159333ac0c2a1536e566d265bdc5809cf3
3
+ size 800325163
data/test-00038.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b1c573aa012299de114c029d3fd649130aad9d8a6b9cb13632cfefb85313632c
3
+ size 707619261
data/test-00039.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:438c8d0b3f633bea64849e7608df5e011ef1c52ba010c92ad753b8d32449c5d7
3
+ size 712257773
data/test-00040.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e1a2ba45104f25ce53bd9591192f6f2a9259d12d807821858944c2c2123af24f
3
+ size 714531504
data/test-00041.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:038cf3a9b39ec5e295ad6880a045135ec8957b4933327a4fe04a847722cc4381
3
+ size 715431141
data/test-00042.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9272aca1294479bf47370388876698060dd4411290723aa78d6ef9f2eeb5949d
3
+ size 719548802