shuaishuaicdp/UI-Video-Removal-Benchmark

UI Video Removal Benchmark

This repository contains a frozen 200-video benchmark for evaluating gameplay UI/HUD removal:

• 100 synthetic paired examples with an input video, clean reference video, mask, synthetic-overlay annotation, and per-frame UI bounding boxes.
• 100 in-the-wild examples with human-reviewed UI trajectory annotations.

All videos are five seconds long, 1280x720, 30 FPS, and 150 frames. Benchmark IDs are frozen as synthetic_000001 through synthetic_000100 and wild_000001 through wild_000100.

Layout

data/
  synthetic/
    input/                 # videos with synthetic HUD overlays
    clean/                 # paired clean videos
    masks/                 # rendered HUD masks
    overlay_annotations/   # synthetic rendering metadata
    per_frame_bboxes/      # frame-level UI bounding boxes
  in_the_wild/
    input/                 # real gameplay videos with native UI
    annotations/           # human-reviewed UI trajectories
manifests/
  synthetic.jsonl
  in_the_wild.jsonl
  all.jsonl
  id_map.jsonl
  legacy_to_new.json

Each manifest row retains the source metadata, original identifiers, curation provenance, media geometry, and annotation format. id_map.jsonl and legacy_to_new.json map historical benchmark IDs to the frozen IDs so earlier in-the-wild model outputs can be reused without ambiguous filename matching.

Evaluation protocol

Model outputs are evaluated after aspect-preserving direct resize to 1280x720. The benchmark itself stores the original 1280x720 media and does not include model-generated outputs or scores.

Provenance and license

The benchmark consolidates curated subsets from the public HUD-Video dataset and the gameplay-benchmark-strict annotation release. Upstream media and annotations retain their original provenance metadata. Users are responsible for complying with upstream media licenses and terms.