Deanshy1/RenderMatte-Animal-Forest-2K

RenderMatte Animal Forest 2K

A synthetic video-matting dataset. Rigged 3D animals are rendered in Blender (Cycles) and composited onto real forest footage, and every frame comes with its 16-bit alpha matte.

Research use only. The animal models and background videos come from the RenderMatte source-asset collection, and not every source-asset license has been reviewed yet. You may use the data for non-commercial research, but do not redistribute it or use it commercially (see LICENSE).

At a glance

Clips	189 = 27 animals × 7 shots (4 full-body + 3 half-body)
Backgrounds	64 real forest video segments. Each segment is used at most 6 times, and no animal reuses a segment.
Length	7 s per clip at the background's own frame rate: 168 frames (24 / 23.976 fps), 175 frames (25 fps) or 210 frames (29.97 / 30 fps)
Frames	36,750 composites + 36,750 mattes
Resolution	1920 × 1080
Size	about 114 GB (composites 109 GB, mattes 3.8 GB, review crops 0.9 GB)

Layout

One folder per clip, named <animal>__<shot>__v<background id>, for example red_fox_3_3__front_full__v174.

<clip>/
├── composite/frame_0001.png …   1920×1080 RGB, 8-bit: the animal composited onto the background
├── mask/mask_0001.png …         1920×1080 single channel, 16-bit: alpha matte (0 = background, 65535 = animal)
├── check/                       three quick-look crops used for manual review (not needed for training)
├── _composite_meta.json         per-frame placement, scale, camera-follow and colour settings
└── _qc.json                     automatic quality-check result (every clip passes)

composite/frame_NNNN.png and mask/mask_NNNN.png with the same number belong together.

Shots:

• Full-body: side_full, front_full, back_low, back_high, back_away_high, diag_high, diag_arc, side_low_arc, top_down
• Half-body: front_half, side_half, rear_half, diag_half, side_low_half

import cv2, numpy as np
clip = "red_fox_3_3__front_full__v174"
rgb   = cv2.imread(f"{clip}/composite/frame_0001.png")[..., ::-1]                         # H×W×3 uint8, RGB
alpha = cv2.imread(f"{clip}/mask/mask_0001.png", cv2.IMREAD_UNCHANGED) / 65535.0        # H×W float in [0, 1]

How it was made

• Rendering: Blender 5.2.1, Cycles, 32 samples per pixel, filter width 1.5, transparent background. Fur and hair are rendered, with hair dynamics switched off so strands don't streak.
• Alpha: taken directly from the renderer, with no keying, thresholding, feathering or erosion. No shadows and no occlusion: the animal is always the top layer.
• Framing: full-body shots keep the whole animal inside the frame at one fixed scale per clip. Half-body shots are rendered at the final 1920×1080 and pasted 1:1, so the cut through the body always lies on the frame edge.
• Camera motion: the background's camera motion is measured. Full-body ground animals move with the ground so they never slide. Flying animals and half-body shots don't follow the background (half-body reads as a tracking shot).
• Placement: ground animals stand in the lower ground zone of the frame and keep their place on the ground for the whole clip; walking and running cycles are performed in place, so the feet stay locked to the ground. Flying animals are placed in the sky and still cross the frame.
• Lighting: the key and fill lights are set per clip from the background's measured brightness (10th/50th/90th luminance percentiles) and a damped version of its colour cast, so the animal sits in the scene's light instead of one fixed rig.
• Motion sampling: poses are sampled at fractional source frames (interpolated), so source animations whose length differs from the clip length play smoothly instead of repeating or skipping frames.
• Backgrounds: screened by eye and with a person/animal detector to exclude real people and animals.

Quality checks

Every clip passed 8 automatic checks:

1. frame counts;
2. subject size;
3. full body not cut by the frame;
4. no green or magenta colour cast;
5. soft (fur) edges present;
6. true 16-bit mattes;
7. camera-follow error ≤ 3 px;
8. the animal is not clipped by the render frame.

The whole set was also checked for stray non-animal objects in the mattes; none were found.

One clip, Rhinopithecus_Somers__front_half__v050, passes under a documented, manually approved exception: the subject covers 54% of the frame against a 55% threshold, because the somersault forces a wider fixed half-body shot. Note also that the somersault animation ends with the monkey seated and still, so most of those 7 clips are static.

Animals

Animal	Folder prefix	Type	Clips
African elephant (walk / stop)	JW0K309A0_African_el	ground	7
Argali, male (Ovis ammon)	asset_2	ground	7
Bactrian camel (grazing)	JU0L506A1_BactrianCa	ground	7
Black jaguar (run)	JW0P2700_BlackJaguar	ground	7
Brown bear, female (run)	UrsusArctos_Female_R	ground	7
Cattle (looking around)	JU0L419A0_Cattle_Loo	ground	7
Egret	Egret_Animated_VFX_G	flying (aerial backgrounds)	7
Elk, male (grazing)	JF0L811A4_MaleElk_Ea	ground	7
Giant panda (trot)	JF0N2N0K_Panda_Trot	ground	7
Giraffe (walk)	Giraffe_Walk_2_83_5	ground	7
Golden snub-nosed monkey (run)	Rhinopithecus_Run_2	ground	7
Golden snub-nosed monkey (somersault)	Rhinopithecus_Somers	ground	7
Grey tabby cat (run)	JC0L219A2_GreyTabbyC	ground	7
Jaguar (run)	JW0N8300_PantheraOnc	ground	7
Labrador retriever (walk)	Dog_Labrador_walking	ground	7
Lion, adult (walk)	JF0LA12A9_Lion_Walk	ground	7
Lion, young (trot)	JF0M111A0_Lion_Young	ground	7
Polar bear (turning)	JF0O8Z0B_PolarBear_3	ground	7
Red fox (walk)	red_fox_3_3	ground	7
Red squirrel (run)	JF0L811A5_RedSquirre	ground	7
Red-crowned crane (fly)	JF0LA12A6_Red_crowne	flying (aerial backgrounds)	7
Red-crowned crane (fly, second model)	JF0LA12A6_Red_crowne	flying (aerial backgrounds)	7
Siberian tiger (lying)	JF0LA12A0_SiberianTi	ground	7
Silverback gorilla (chest beat)	JF0N2N0R_SilverbackG	ground	7
Steppe eagle (fly)	JF0L810A0_AquilaNipa	flying (aerial backgrounds)	7
White horse (run)	JW0N4600_Horse_White	ground	7
Wolf (howl)	JF0N2N0G_Wolf_Howl_3	ground	7

Revision history

v2 (2026-10). All 189 clips re-rendered after a review of v1. Animals, shots, backgrounds and time windows are unchanged; four systematic problems were fixed:

• Foot sliding. In v1 a "walking" animal was moved across a fixed 39% of the frame width regardless of its real speed, while its legs kept their own rhythm, so the feet slid over the ground. Measured gait speeds showed that correct speed, a 7-second clip, a large on-screen animal and staying inside the frame cannot all hold at once (a jaguar at 3.6 m/s covers 25 m in 7 s). v2 keeps ground animals in place on the ground; flying animals still cross the frame.
• Stutter. v1 rounded each output frame to a whole source frame, so 167 of 189 clips repeated or skipped source frames unevenly, and the composite placed the animal at whole-pixel positions, so slow drift moved in one-pixel jumps. v2 interpolates poses at fractional source frames and places the animal with sub-pixel precision. A slow vertical "bob" unrelated to the gait was removed.
• Lighting. v1 lit every clip with the same sun and a green-tinted fill. v2 derives the light from each background (see How it was made).
• One clip re-framed. JW0N4600_Horse_White__side_full__v049 sits on a background whose camera pans by about 1200 px; the galloping horse is shown smaller (about 37% of the frame height instead of 51%) so it stays inside the frame while following the ground.