CrossVideoReasoning commited on
Commit
aa2aa7d
·
verified ·
1 Parent(s): 78b3f3b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +36 -19
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`, `kubric_data.tar`, `clvr_data.tar` | 4,827 unique clips | Rendered videos referenced by the questions |
 
45
 
46
- Evaluation and training draw on **disjoint video pools** (e.g. CLEVRER's validation split for evaluation,
47
- its training split for training).
 
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
- ### Expected layout
62
 
63
- Video paths in the annotations are relative to the repository root and resolve to a `videos/` directory:
64
 
65
  ```
66
  videos/clvr/video_validation/video_11194.mp4
67
- videos/habitat_data/...
 
68
  videos/kubric_data/scene_144/cam2.mp4
69
  ```
70
 
71
- Extract the archives into `videos/` before running evaluation:
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
- | Entity Resolution | Object Re-identification | Habitat | 2 | 25% |
86
- | Entity Resolution | Spatial Measurement | Kubric | 2 | 25% |
87
  | Comparative Reasoning | Numerical Comparison | CLEVRER | 2 | **20%** |
88
  | Comparative Reasoning | Kinematic Comparison | CLEVRER | 2 | 25% |
89
- | Holistic Aggregation | Object Counting | Habitat | 3 | 25% |
90
- | Holistic Aggregation | Route Planning | Habitat | 3 | 25% |
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. Rendered clips remain subject to the
141
- licenses of their upstream simulators and scene assets; consult those before redistribution.
 
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**.