MonsterDie commited on
Commit
c71274f
·
verified ·
1 Parent(s): acf19f1

Document dataset annotation methodology

Browse files
Files changed (1) hide show
  1. README.md +18 -0
README.md CHANGED
@@ -10,6 +10,24 @@ The dataset covers **12,079 scenes from 3,188 videos across 9 source channels**.
10
 
11
  The data processing pipeline is maintained in a separate repository. [SpatialSemanticPipeline](https://github.com/monsterdie12138/SpatialSemanticPipeline)
12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
13
  ## Repository layout
14
 
15
  | Path | Contents |
 
10
 
11
  The data processing pipeline is maintained in a separate repository. [SpatialSemanticPipeline](https://github.com/monsterdie12138/SpatialSemanticPipeline)
12
 
13
+ ## Annotation methodology
14
+
15
+ The accompanying pipeline uses TransNetV2 to detect scene boundaries in
16
+ source videos. It runs VGGT-Omega on overlapping video chunks to estimate
17
+ camera positions and orientations. The pipeline checks trajectory quality,
18
+ and the released pose CSVs include a `status` field. Rows marked `bad`
19
+ can contain missing pose values.
20
+
21
+ A vision-language model generates timestamped English annotations from
22
+ video clips. When trajectory-derived navigation hints are available, the
23
+ pipeline includes them in the prompt. Otherwise, it uses a prompt without
24
+ hints. Each annotation contains an `Observation` describing visible
25
+ content, a `Latent Thought` expressing the model's inference about intent
26
+ or reasoning, and an `Action` describing movement.
27
+
28
+ See [SpatialSemanticPipeline](https://github.com/monsterdie12138/SpatialSemanticPipeline)
29
+ for the implementation, configuration, and prompt templates.
30
+
31
  ## Repository layout
32
 
33
  | Path | Contents |