Voxel51/sama_material_centric_video_dataset

Dataset Card for sama_material_aware

This is the training and evaluation dataset introduced alongside SAMa: Material-Aware 3D Selection and Segmentation. It is an object-centric synthetic video dataset with dense per-frame, per-material pixel-level segmentation annotations, designed to fine-tune video object-selection models for the task of material selection.

This FiftyOne dataset contains 500 samples (450 train, 50 test).

Installation

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

dataset = load_from_hub("Voxel51/sama_material_centric_video_dataset")

session = fo.launch_app(dataset)

---

Dataset Details

Dataset Description

The SAMa dataset is a synthetic, procedurally generated video dataset of 3D objects rendered across 30-frame sequences with multiple camera trajectories. Each video depicts one or more 3D objects, each composed of several distinct surface materials. The dataset provides dense per-pixel material segmentation masks for every frame, encoded in COCO RLE format.

The key motivation for the video format (rather than static images) is multiview consistency: by training on temporally coherent material annotations, a video model learns to produce consistent material selections across viewpoints — a prerequisite for lifting 2D material selection into 3D representations (NeRFs, 3D Gaussians, meshes).

• Curated by: Michael Fischer (UCL / Adobe Research), Iliyan Georgiev, Thibault Groueix, Vladimir G. Kim, Valentin Deschaintre (Adobe Research), Tobias Ritschel (UCL)
• Language(s): N/A (vision dataset)
• License: [More Information Needed]

Dataset Sources

• Paper: SAMa: Material-Aware 3D Selection and Segmentation
• Project page / Demo: https://mfischer-ucl.github.io/sama/

---

Uses

Direct Use

• Fine-tuning video segmentation models (e.g. SAM2) for per-material selection
• Training and evaluating material segmentation models that require cross-frame consistency
• Benchmarking material-aware 3D selection methods on synthetic object-centric scenes

Out-of-Scope Use

• General semantic segmentation (labels are appearance-based material instances, not semantic categories like "wood" or "metal")
• Real-world scene understanding — all content is synthetic
• Glass, mirror, or highly transparent materials are acknowledged limitations of the method trained on this data

---

Dataset Structure

Splits

Split	Videos	Frames
train	450	13,500
test	50	1,500

FiftyOne Schema

Each sample in the dataset corresponds to one video and has the following fields:

Sample-level fields:

Field	Type	Description
filepath	string	Path to the H.264 MP4 video (1024×1024, 30 frames, 10 FPS)
split	string	"train" or "test"
obj_idx	int	Object index — matches the integer N in video{N}.mp4 / annotations{N}.json
render_mode	string	Camera trajectory: "flyover", "turntable", "flyovernoLookatZoomOut", etc.
n_materials	int	Number of distinct material regions in this video

Frame-level fields (accessed via sample.frames[1] through sample.frames[30], 1-indexed):

Field	Type	Description
segmentation	fo.Segmentation	Per-pixel material index mask stored as a uint8 PNG on disk

The segmentation mask is a uint8 grayscale PNG where:

• Pixel value 0 = background
• Pixel value 1 through N = material index (1-based), corresponding to dataset.mask_targets["segmentation"]

Dataset-level:

• dataset.mask_targets["segmentation"] maps integer pixel values to generic label strings: {1: "material_1", 2: "material_2", ...} up to the maximum number of materials observed. Human-readable material names are not included in this release.
• dataset.info contains source and paper metadata.

Parsing Notes

• Original videos are encoded as MPEG-4 Part 2 and must be re-encoded to H.264 for FiftyOne's App to stream them. The reencode_h264.sh script handles this.
• Annotations in the raw release are COCO RLE JSON files (annotations{N}.json). The build_fiftyone_dataset.py script decodes them using pycocotools and saves per-frame uint8 PNG mask files under masks/.
• FiftyOne uses 1-based frame indexing: annotation frame index 0 maps to sample.frames[1].
• Video frames are rendered in linear RGB space. For visualization, apply sRGB gamma correction (frame ** 0.4545).

---

Dataset Creation

Curation Rationale

Prior material segmentation datasets (Materialistic, OpenSurfaces, Dense Material Segmentation) only contain static images, which cannot teach multiview consistency — a critical requirement for lifting 2D material selection into 3D. This dataset was purpose-built to address that gap by providing temporally consistent per-material annotations across rendered camera trajectories.

Source Data

Data Collection and Processing

Videos were procedurally generated in Blender 4.3 using the Principled BSDF shader:

• 3D objects: Randomly sampled from a subset of 9,082 textured Substance 3D objects. Objects with only a single material were excluded (no supervision signal for selection).
• Materials: Randomly sampled from a library of 29,472 Substance material maps (including multiple variants of the same base material, e.g., different wood types). Each material is assigned to appear at least twice per scene to provide positive supervision and disambiguate material selection from object selection.
• Scene layout: Each video contains at least two objects, displaced by up to half their bounding box extent to reduce spatial overlap.
• Lighting: Randomly selected HDR environment maps from a set of 420 HDRIs sourced from PolyHaven. Objects are alpha-composited onto the environment map background to reduce the domain gap between synthetic and real-world data.
• Camera trajectories: Four motion patterns used at specific sampling rates:
  • Turntable (spherical, fixed-radius): 33%
  • Flyover: 33%
  • Zoom-in: 22%
  • Zoom-out: 12%

Who are the source data producers?

The dataset was created by the SAMa paper authors at Adobe Research and University College London. 3D assets are from Adobe's Substance 3D library; HDRIs are from PolyHaven (CC0 licensed).

Annotations

Annotation process

Annotations are automatically generated as part of the Blender rendering pipeline. Dense per-pixel material indices are rendered directly from the scene graph using integer material IDs, with label 0 reserved for the background. No human annotators were involved. Annotations are stored as COCO RLE-encoded binary masks, one per material per frame.

Who are the annotators?

Automated (procedural rendering in Blender 4.3).

Personal and Sensitive Information

None. The dataset contains only synthetic 3D renders of objects and materials.

---

Bias, Risks, and Limitations

• Synthetic domain gap: All content is synthetically rendered. Performance may degrade on real-world captures with complex lighting, subsurface scattering, or non-Principled BSDF materials.
• Transparent and mirror-like materials: The paper explicitly identifies glass and mirrors as failure cases — it is unclear whether a user would prefer to select the transparent/mirror material or what is behind/reflected.
• Fixed material definition: Materials are defined by shared texture and reflectance properties. The model cannot adapt to user-defined selection criteria (e.g., selecting only by roughness).
• Depth dependency: The 3D lifting pipeline depends on accurate depth estimation. Errors in volumetric reconstruction can cause noise in the similarity point cloud.

Recommendations

Users should be aware that this dataset targets appearance-based material similarity, not semantic categories. Two visually similar materials of different types (e.g., two different metals with the same texture) will be treated as the same material; two visually distinct materials of the same category (e.g., two different woods) will be treated as different.

---

Citation

BibTeX:

@article{fischer2024sama,
  title={SAMa: Material-Aware 3D Selection and Segmentation},
  author={Fischer, Michael and Georgiev, Iliyan and Groueix, Thibault and Kim, Vladimir G. and Ritschel, Tobias and Deschaintre, Valentin},
  journal={arXiv preprint arXiv:2411.19322},
  year={2024}
}