robbyant/lingbot-video-rbench

RBench Evaluation Notes

This directory contains the videos and structured captions generated by the LingBot-Video model on the ti2v (text/image-to-video) task against the RBench benchmark. The outputs can be organized into the official RBench evaluation layout via organize_videos_rbench.py and then submitted for scoring.

1. Directory Structure

 rbench/                              # one subdirectory per test_case
├── test_case_rbench_emb_dual_arm_0000_dual_arm_0001/
│   ├── caption.json
│   ├── generate_seed_10.mp4
│   ├── compare_seed_10.mp4
│   └── generate_seed_10_eval_16fps.mp4
├── ...                               # 650 test_case subdirectories in total
└── README.md                         # this file
├── organize_videos_rbench.py         # script that reorganizes this directory into the RBench eval layout

---

2. test_case Naming Convention

Each subdirectory is named as:

test_case_rbench_<category>_<NNNN>_<category_human>_<NNNN+1>

Example: test_case_rbench_emb_dual_arm_0003_dual_arm_0004

• Category segment: emb_dual_arm
• Case index: 0003 (the script uses the second 4-digit number as the final video id 0004; see Section 7)

The second segment dual_arm is the human-readable form of the category (some categories contain hyphens, e.g. long-horizon_planning, multi-entity_collaboration).

---

3. Categories and Counts

There are 650 test_cases in total, distributed as follows:

Group	Category	Count
emb (embodied / robotics)	emb_single_arm	100
emb	emb_dual_arm	100
emb	emb_humanoid	100
emb	emb_quad	100
rbench (reasoning / planning)	rbench_common_manipulation	50
rbench	rbench_long_horizon_planning	50
rbench	rbench_multi_entity_collaboration	50
rbench	rbench_spatial_relationship	50
rbench	rbench_visual_reasoning	50
	Total	650

---

4. File Composition of a Single Case

Under normal conditions each test_case contains 4 files:

File	Description
caption.json	Structured caption for this case (see Section 5)
generate_seed_10.mp4	Raw model-generated video (seed=10); the video used for evaluation
compare_seed_10.mp4	Spliced / reference video for comparison (contains GT or multi-model side-by-side)
generate_seed_10_eval_16fps.mp4	Evaluation version of the generated video re-encoded to 16fps

---

5. caption.json Format

All files have been unified into the outer {"cap": <string>} form: the value of cap is a stringified structured JSON (compact, default separators , / : ), with the outer object indented by 4 spaces.

{
    "cap": "{\"comprehensive_description\": {...}, \"camera_info\": {...}, ...}"
}

It requires a two-step parse when reading: inner = json.loads(obj["cap"]). Schema of inner:

Top-level field	Type	Meaning
comprehensive_description	object	Overall description of scene and camera motion
├ scene_content_description	string	Natural-language description of scene content (objects, actions, temporal order)
└ camera_movement_description	string	Camera motion description (usually static / fixed angle)
camera_info	object	Camera / framing attributes: color / frame_size / shot_type_angle / lens_size / composition / lighting / lighting_type
world_knowledge	list	World knowledge / priors, usually empty []
prominent_elements	list	List of prominent elements (objects) in the frame, see below

---

6. Video File Notes

• The main video for evaluation is always generate_seed_10.mp4 (seed=10).
• compare_seed_10.mp4 is for visual comparison only and is not scored.
• All videos are mp4/H.264, ~5s long, 24fps (the eval variant is 16fps).

---

7. RBench Evaluation Preparation and Submission Flow

The original layout in this directory is not yet the official RBench evaluation layout. Going from rbench.zip to a completed official evaluation takes 4 steps:

Step 1: Unzip rbench.zip

unzip rbench.zip -d ./rbench     # the zip uses a flat one-level structure; unzipping yields the 650 test_case_rbench_*/ directories and rebench_video.zip

If ./rbench already exists (i.e. this directory itself is the unzipped output), this step can be skipped.

Step 2: Organize into the RBench Evaluation Layout

python organize_videos_rbench.py

Script logic (key points):

• Source directory: ./rbench; output directory: ./ready_rbench

• For each test_case it takes generate_seed_10.mp4 (skips with a warning if missing)

• It maps the category to an RBench directory name according to the table below, and uses the case index (the second 4-digit number in the name) as the file name:

  Original category segment	RBench directory
  emb_single_arm	single_arm
  emb_dual_arm	dual_arm
  emb_humanoid	humanoid
  emb_quad	quad
  rbench_common_manipulation	common_manipulation
  rbench_long_horizon_planning	long-horizon_planning
  rbench_multi_entity_collaboration	multi-entity_collaboration
  rbench_spatial_relationship	spatial_relationship
  rbench_visual_reasoning	visual_reasoning

• Final output layout (required by RBench evaluation):

  ready_rbench/
  └── <category>/
      └── videos/
          └── <NNNN>.mp4     # e.g. 0004.mp4
  

• If ./ready_rbench already exists, the script does not clear it automatically; it prints a WARNING and you need to remove it manually before re-running.

• Existing target files are skipped (prints EXISTS (skip)).

Step 3: Clone the Official RBench Repository

git clone https://github.com/DAGroup-PKU/ReVidgen.git

ReVidgen is the official codebase for RBench evaluation, containing the scoring scripts and dependency notes. After cloning, follow its README.md to install dependencies (recommended inside an isolated venv to avoid polluting the main environment).

Step 4: Run the Official Evaluation

Following the README.md of the ReVidgen repository, pass the ./ready_rbench/ output from Step 2 as the input directory to its evaluation script and run the official RBench scoring. The exact command and arguments are subject to the official repository's documentation (they may differ across versions), for example:

cd ReVidgen
# After installing dependencies per the official README, run the eval entry point with the input pointing to ../ready_rbench
# python <official_eval_script> --video_dir ../ready_rbench ...