Deanshy1/RenderMatte-Human-Multi-Street-2K

RenderMatte Human Multi-Person Street 2K

A synthetic video-matting dataset for the multi-person case. Two or three rigged 3D human characters are animated and rendered together in one Blender (Cycles) scene, then composited onto real street footage. Every frame comes with its 16-bit alpha matte.

It is built on top of the single-person VideoMatting pipeline and its source-asset collection, and keeps the same render and compositing settings wherever possible.

Research use only. The character models, motions and background videos come from the VideoMatting source-asset collection, and not every source-asset license has been reviewed. Use the data for non-commercial research only; do not redistribute it or use it commercially (see LICENSE).

At a glance

Clips	300
Split	train 255 / test 45. No character and no background segment is shared between the two splits.
Composition	120 × three people full-body, 120 × two people full-body, 60 × two people half-body
Length	6 s / 7 s / 8 s at 30 fps → 180 / 210 / 240 frames (101 / 95 / 104 clips)
Frames	63,090 composites + 63,090 mattes
Resolution	1920 × 1080
Characters	147 distinct characters, 720 appearances, a single character appears at most 6 times
Backgrounds	127 real video segments, each used at most 4 times, with staggered start times. Where a segment is too short for four separate windows the windows overlap: of the 267 clip pairs that share a segment, 230 overlap in time (median overlap 86% of the shorter clip, 29 pairs almost fully). Such clips are therefore not independent samples, although their characters, motions and placement always differ. Train and test share no segment at all.
Scenes	outdoor daytime scenes, predominantly streets and pedestrian streets. The segments come from three stock collections: 124 segments / 291 clips from a general urban-and-landscape collection, 2 segments / 6 clips forest-themed, 1 segment / 3 clips an indoor venue.
Size	35.4 GB total: composites 30.3 GB, mattes 4.5 GB, preview videos 0.7 GB, metadata 10 MB

Layout

One folder per clip. The folder name encodes the split and the composition: g3f = three people full-body, g2f = two people full-body, g2h = two people half-body.

<train|test>/<clip>/                  e.g. train/train_g3f_0014/
├── composite/00001.jpg …             1920×1080 RGB: the characters composited onto the background
├── mask/00001.png …                  1920×1080 single channel, 16-bit: alpha matte
├── video/preview_1920x1080.mp4       H.264 preview of this clip
├── record.json                       characters, background segment and time window, motion presets,
│                                     assigned heights, placement, QC result
└── group_report.json                 camera, lighting, final measured height of each character,
                                      seconds per frame, render backend

composite/NNNNN.jpg and mask/NNNNN.png with the same number belong together. Numbering starts at 00001 and is contiguous within a clip.

import cv2
clip = "train/train_g3f_0014"
rgb   = cv2.imread(f"{clip}/composite/00001.jpg")[..., ::-1]                        # H×W×3 uint8, RGB
alpha = cv2.imread(f"{clip}/mask/00001.png", cv2.IMREAD_UNCHANGED) / 65535.0        # H×W float in [0, 1]

Conventions

• One matte per group, not per person. All characters in a clip are merged into a single alpha channel (0 = background, 65535 = character, soft values on edges and hair). Supervising "matte only person 2" is therefore not possible with this release.
• Composites are JPEG (-q:v 3 -pix_fmt yuvj444p, no chroma subsampling) — the same setting the single-person pipeline uses. The exact foreground and background layers cannot be recovered from them. The mattes are lossless 16-bit PNG.
• Character size is uniform, not matched to the background. In full-body clips a character spans about half the frame height with the feet at 90% of the frame height; in half-body clips the crop falls near the lower legs and the character nearly fills the frame height. Character size is not adapted to the apparent size of real pedestrians in the background.
• Heights are normalised to 1.60–1.85 m. group_report.json → heights_m is the height actually applied (verified to lie in [1.600, 1.850] for all 720 appearances). Note that record.json → character_height is the source model's own measurement and is not in consistent units — some models report centimetres (e.g. 185.9) or arbitrary units (e.g. 1.12, 4.95). Use heights_m.
• Characters do not overlap each other. They are spread left to right and checked frame by frame; the smallest gap between two characters across the release is 0.14 m.
• Backgrounds contain real people. Most clips have pedestrians, cyclists or vehicles somewhere in frame. Matting is prompt-directed in the intended use case, so this does not affect the label. Real pedestrians are of course not included in the matte.
• No close foreground occluders. For each clip a time window was chosen in which nothing walks or drives in front of the area where the characters stand; static objects (poles, tree trunks, railings, parked cars) were screened per clip by eye.
• Slowly drifting camera. In 208 clips the background camera drifts. The characters and the matte follow the background frame by frame, so they read as standing still. Shaky footage was not used.

How it was made

1. Per character: motion generated and passed through the single-person pipeline's own checks, then cached with a per-frame bounding box. Upstream code unmodified.
2. Grouping: 2–3 characters placed in one scene, heights normalised, motion start times staggered, spread left to right with a per-frame non-overlap check, framed by one shared orthographic camera.
3. Rendering: Blender 5.0.1, Cycles, 64 samples with denoising, transparent background, 16-bit RGBA.
4. Compositing: the matte is the rendered alpha channel; the composite lays the characters over the corresponding time window of the background video.

36 clips were rendered on GPU and 264 on CPU (the GPU queue was slow). The two backends were checked against each other on one group re-rendered both ways (12 frames): matte IoU 0.99999 with no pixel differing by more than 0.1/65535, and composite PSNR of at least 56.9 dB with a mean difference of 0.57/255 on the characters and exactly zero outside them.

Quality checks

Every clip passes all of the following automatically:

• number of connected components in the matte does not exceed the number of characters (catches hair or accessories flying off);
• no character touches the frame border;
• the group's height in frame matches the plan;
• the matte has soft edges (not a hard binary mask);
• composite and matte frame counts are complete and equal.

All 300 clips pass (record.json → qc.verdict == "PASS"). Frames were also spot-checked by eye.

Known limitations

• Characters have no contact shadows and are never occluded by objects in the background — this is why time windows without close foreground occluders were selected. Both apply to the single-person data as well.
• The background is flat video with no real ground relief; all characters' feet land on one line.
• One matte per group, so per-person supervision is not available.
• In half-body clips characters are large, and long-haired characters render slowly (up to 78 s per frame; median across the release is 12.4 s).
• record.json contains absolute paths on the render cluster and the original filenames of the source assets. They are kept for traceability and are not meant to resolve on another machine.