File size: 2,799 Bytes
3992795
 
 
b36205c
3992795
 
 
 
 
 
 
 
 
 
 
6f00042
e1d7b1c
80d7ebf
 
f27dd75
74a2643
e1d7b1c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
74a2643
1b352bb
 
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
---
license: mit
task_categories:
  - video-to-video
tags:
  - physics
  - video-editing
  - benchmark
  - rigid-body
  - simulation
pretty_name: PCVE-RigidBench
size_categories:
  - n<1K
---

# PCVE-RigidBench

[![GitHub](https://img.shields.io/badge/GitHub-PCVE--RigidBench-black?logo=github)](https://github.com/Hammour-steak/PCVE-RigidBench)

Physics Counterfactual Video Editing benchmark. 20 scenes,
20 source videos, 129 edit tasks.

## Layout
- `benchmark_manifest.json`: flat index of every source + edit; the
  authoritative record. Every field the evaluator needs (prompts,
  physics_diff, edit_summary, video/gt paths) is inlined here.
- `scenes/{scene}/cases/{case_id}/`:
    - `video.mp4` -- the source (in the baseline case) or edited render.
    - `prompts.json` -- redundant with the top-level manifest; kept
      per-case so an evaluator that iterates directories has everything
      local.
    - `ground_truth_transforms.json` -- per-frame world matrices,
      linear/angular velocities and any object-specific quality metrics
      the physics sim produced. Use these for physics-consistency
      metrics; skip them if you are only doing pixel/perceptual eval.
- `thumbnails/{scene}/{case_id}.jpg`: mid-frame preview per case.
- `splits/all.txt`: every edit's `global_id`, one per line.
- `splits/by_property/`: same, bucketed by the edited property (mass /
  friction / restitution / initial_velocity / presence for DELETE and ADD
  edits, which both move an object along the present/absent axis).
- `splits/by_timing/`: same, bucketed by when the edit takes effect --
  `whole_clip.txt` for edits that hold from the first frame, and
  `partway.txt` for edits that land at `applies_from_frame`, where the
  frames before it are supposed to match the source video exactly.

## Prompts
Each edit has 4 flavors under `prompts`:
- `vague.zh` / `vague.en`: direction-only ("调大一点" / "Decrease X").
- `quantitative.zh` / `quantitative.en`: includes `from` and `to` numbers.

Every prompt opens by saying when the edit takes effect, because that is
part of the instruction: "From frame 1 onwards" / "From the start of the
clip" for the whole-clip edits, and "From frame 22 onwards" / "Early in
the clip" for one that lands partway through. `applies_from_frame` in the
manifest is the same fact in machine-readable form.

## Task
Given (`source_video`, prompt), generate a video that matches
`edited_video`. Compare against `ground_truth_transforms.json` for
per-frame object trajectories, or against `edited_video` for pixel /
perceptual metrics.

## Counts
- Edits by kind: {'ADD': 7, 'DELETE': 30, 'SET': 92}
- Edits by timing: {'partway': 12, 'whole_clip': 117}
- Edits by property: {'friction': 21, 'initial_velocity': 26, 'mass': 32, 'presence': 37, 'restitution': 13}