Datasets:
Update README.md
Browse files
README.md
CHANGED
|
@@ -41,15 +41,16 @@ from human annotation. The same generators produce both an evaluation benchmark
|
|
| 41 |
|---|---|---|
|
| 42 |
| `test/metadata.jsonl` | **4,000** | Evaluation benchmark, 500 questions for each of 8 tasks |
|
| 43 |
| `train/sft_all8_answer_only.json` | **15,960** | Training set, chat-formatted, answer-only targets |
|
| 44 |
-
| `habitat_data.tar`
|
|
|
|
| 45 |
|
| 46 |
-
|
| 47 |
-
|
|
|
|
| 48 |
|
| 49 |
### Loading
|
| 50 |
|
| 51 |
-
The benchmark and the training set have different schemas, so they are published as
|
| 52 |
-
two configs:
|
| 53 |
|
| 54 |
```python
|
| 55 |
from datasets import load_dataset
|
|
@@ -58,22 +59,39 @@ test = load_dataset("CrossVideoReasoning/SYNCR", split="test") #
|
|
| 58 |
train = load_dataset("CrossVideoReasoning/SYNCR", "training", split="train") # 15,960 examples
|
| 59 |
```
|
| 60 |
|
| 61 |
-
##
|
| 62 |
|
| 63 |
-
Video paths in the annotations are relative to the repository root and resolve
|
| 64 |
|
| 65 |
```
|
| 66 |
videos/clvr/video_validation/video_11194.mp4
|
| 67 |
-
videos/
|
|
|
|
| 68 |
videos/kubric_data/scene_144/cam2.mp4
|
| 69 |
```
|
| 70 |
|
| 71 |
-
|
| 72 |
|
| 73 |
```bash
|
| 74 |
mkdir -p videos && for f in *_data.tar; do tar -xf "$f" -C videos/; done
|
| 75 |
```
|
| 76 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 77 |
## Tasks
|
| 78 |
|
| 79 |
Eight tasks in four reasoning families, 500 evaluation questions each.
|
|
@@ -82,12 +100,12 @@ Eight tasks in four reasoning families, 500 evaluation questions each.
|
|
| 82 |
|---|---|---|---|---|
|
| 83 |
| Temporal Alignment | Multi-Angle Synchronization | Kubric | 3 | 25% |
|
| 84 |
| Temporal Alignment | Sequential Ordering | CLEVRER | 4 | 25% |
|
| 85 |
-
|
|
| 86 |
-
|
|
| 87 |
| Comparative Reasoning | Numerical Comparison | CLEVRER | 2 | **20%** |
|
| 88 |
| Comparative Reasoning | Kinematic Comparison | CLEVRER | 2 | 25% |
|
| 89 |
-
| Holistic
|
| 90 |
-
| Holistic
|
| 91 |
|
| 92 |
Numerical Comparison presents five options; every other task presents four. Correct answers are balanced
|
| 93 |
across option positions (975 each for A–D, plus 100 at E for the five-option task).
|
|
@@ -98,7 +116,7 @@ Each row of `test/metadata.jsonl`:
|
|
| 98 |
|
| 99 |
| Field | Type | Description |
|
| 100 |
|---|---|---|
|
| 101 |
-
| `category` | string | Reasoning family |
|
| 102 |
| `task` | string | Task name |
|
| 103 |
| `question` | string | The question text |
|
| 104 |
| `options` | list[string] | Answer choices, each prefixed `A)`, `B)`, … |
|
|
@@ -133,12 +151,13 @@ the `video` field carries the same relative paths as the benchmark.
|
|
| 133 |
|
| 134 |
Videos are rendered from three simulators, each with its own upstream terms:
|
| 135 |
|
| 136 |
-
- **CLEVRER** — Yi et al., ICLR 2020
|
| 137 |
- **Kubric** — Greff et al., CVPR 2022
|
| 138 |
- **Habitat** — Savva et al., ICCV 2019 / Szot et al., NeurIPS 2021
|
| 139 |
|
| 140 |
-
The annotations in this repository are released under CC BY 4.0.
|
| 141 |
-
|
|
|
|
| 142 |
|
| 143 |
## Changes from the previous release
|
| 144 |
|
|
@@ -146,8 +165,6 @@ This release supersedes the earlier version of this dataset and is **not** a dro
|
|
| 146 |
|
| 147 |
- The evaluation set is now **4,000 questions** (500 per task), replacing the previous 8,163-example
|
| 148 |
single split, and a **15,960-example training split** has been added.
|
| 149 |
-
- Two families were renamed: *Spatial Tracking* → **Entity Resolution**, *Holistic Synthesis* →
|
| 150 |
-
**Holistic Aggregation**.
|
| 151 |
- Tasks were renamed and regenerated: *CLEVRER collision comparison* → **Numerical Comparison**,
|
| 152 |
*Kubric spatial tracking* → **Spatial Measurement**, *Habitat agent/object tracking* →
|
| 153 |
**Object Re-identification**.
|
|
|
|
| 41 |
|---|---|---|
|
| 42 |
| `test/metadata.jsonl` | **4,000** | Evaluation benchmark, 500 questions for each of 8 tasks |
|
| 43 |
| `train/sft_all8_answer_only.json` | **15,960** | Training set, chat-formatted, answer-only targets |
|
| 44 |
+
| `habitat_data.tar` | — | Habitat clips used by both splits |
|
| 45 |
+
| `kubric_data.tar` | — | Kubric clips used by both splits |
|
| 46 |
|
| 47 |
+
The evaluation set draws on **4,827** distinct video files and the training set on **14,956**, with **no
|
| 48 |
+
video shared between the two splits**. (These are unions over tasks; per-task counts sum to more, because
|
| 49 |
+
some videos are used by more than one task.)
|
| 50 |
|
| 51 |
### Loading
|
| 52 |
|
| 53 |
+
The benchmark and the training set have different schemas, so they are published as two configs:
|
|
|
|
| 54 |
|
| 55 |
```python
|
| 56 |
from datasets import load_dataset
|
|
|
|
| 59 |
train = load_dataset("CrossVideoReasoning/SYNCR", "training", split="train") # 15,960 examples
|
| 60 |
```
|
| 61 |
|
| 62 |
+
## Getting the videos
|
| 63 |
|
| 64 |
+
Video paths in the annotations are relative to the repository root and resolve under `videos/`:
|
| 65 |
|
| 66 |
```
|
| 67 |
videos/clvr/video_validation/video_11194.mp4
|
| 68 |
+
videos/clvr/video_train/video_05825.mp4
|
| 69 |
+
videos/habitat_data/route_plan/...
|
| 70 |
videos/kubric_data/scene_144/cam2.mp4
|
| 71 |
```
|
| 72 |
|
| 73 |
+
**1. Habitat and Kubric** — extract the archives in this repository:
|
| 74 |
|
| 75 |
```bash
|
| 76 |
mkdir -p videos && for f in *_data.tar; do tar -xf "$f" -C videos/; done
|
| 77 |
```
|
| 78 |
|
| 79 |
+
**2. CLEVRER** — these clips are **not redistributed here**. Download them from the official CLEVRER
|
| 80 |
+
release at **http://clevrer.csail.mit.edu** and place them so the paths above resolve:
|
| 81 |
+
|
| 82 |
+
```bash
|
| 83 |
+
mkdir -p videos/clvr
|
| 84 |
+
unzip video_validation.zip -d videos/clvr/ # 1,987 clips used by the benchmark
|
| 85 |
+
unzip video_train.zip -d videos/clvr/ # 5,879 clips used by the training set
|
| 86 |
+
```
|
| 87 |
+
|
| 88 |
+
Which sources each split needs:
|
| 89 |
+
|
| 90 |
+
| Split | CLEVRER | Habitat | Kubric | Total distinct |
|
| 91 |
+
|---|---|---|---|---|
|
| 92 |
+
| Evaluation | 1,987 (`video_validation`) | 1,086 | 1,754 | **4,827** |
|
| 93 |
+
| Training | 5,879 (`video_train`) | 4,958 | 4,119 | **14,956** |
|
| 94 |
+
|
| 95 |
## Tasks
|
| 96 |
|
| 97 |
Eight tasks in four reasoning families, 500 evaluation questions each.
|
|
|
|
| 100 |
|---|---|---|---|---|
|
| 101 |
| Temporal Alignment | Multi-Angle Synchronization | Kubric | 3 | 25% |
|
| 102 |
| Temporal Alignment | Sequential Ordering | CLEVRER | 4 | 25% |
|
| 103 |
+
| Spatial Tracking | Object Re-identification | Habitat | 2 | 25% |
|
| 104 |
+
| Spatial Tracking | Spatial Measurement | Kubric | 2 | 25% |
|
| 105 |
| Comparative Reasoning | Numerical Comparison | CLEVRER | 2 | **20%** |
|
| 106 |
| Comparative Reasoning | Kinematic Comparison | CLEVRER | 2 | 25% |
|
| 107 |
+
| Holistic Synthesis | Object Counting | Habitat | 3 | 25% |
|
| 108 |
+
| Holistic Synthesis | Route Planning | Habitat | 3 | 25% |
|
| 109 |
|
| 110 |
Numerical Comparison presents five options; every other task presents four. Correct answers are balanced
|
| 111 |
across option positions (975 each for A–D, plus 100 at E for the five-option task).
|
|
|
|
| 116 |
|
| 117 |
| Field | Type | Description |
|
| 118 |
|---|---|---|
|
| 119 |
+
| `category` | string | Reasoning family (`Temporal_Alignment`, `Spatial_Tracking`, `Comparative_Reasoning`, `Holistic_Synthesis`) |
|
| 120 |
| `task` | string | Task name |
|
| 121 |
| `question` | string | The question text |
|
| 122 |
| `options` | list[string] | Answer choices, each prefixed `A)`, `B)`, … |
|
|
|
|
| 151 |
|
| 152 |
Videos are rendered from three simulators, each with its own upstream terms:
|
| 153 |
|
| 154 |
+
- **CLEVRER** — Yi et al., ICLR 2020 — http://clevrer.csail.mit.edu
|
| 155 |
- **Kubric** — Greff et al., CVPR 2022
|
| 156 |
- **Habitat** — Savva et al., ICCV 2019 / Szot et al., NeurIPS 2021
|
| 157 |
|
| 158 |
+
The annotations in this repository are released under CC BY 4.0. CLEVRER clips are not redistributed here
|
| 159 |
+
and remain under their original terms. Rendered Habitat and Kubric clips remain subject to the licenses of
|
| 160 |
+
those simulators and their scene assets; consult those before redistribution.
|
| 161 |
|
| 162 |
## Changes from the previous release
|
| 163 |
|
|
|
|
| 165 |
|
| 166 |
- The evaluation set is now **4,000 questions** (500 per task), replacing the previous 8,163-example
|
| 167 |
single split, and a **15,960-example training split** has been added.
|
|
|
|
|
|
|
| 168 |
- Tasks were renamed and regenerated: *CLEVRER collision comparison* → **Numerical Comparison**,
|
| 169 |
*Kubric spatial tracking* → **Spatial Measurement**, *Habitat agent/object tracking* →
|
| 170 |
**Object Re-identification**.
|