allenai/WildDet3D-visualization-source

WildDet3D Visualization Data

This repository hosts the visualization data for the WildDet3D-Bench benchmark — a human-annotated evaluation set for monocular 3D object detection in the wild.

Dataset Overview

WildDet3D-Bench is a validation set of 2,470 images drawn from three source datasets, with 9,256 human-verified 3D bounding box annotations across 2,196 images.

Source	Images	Description
COCO Val	424	MS-COCO 2017 validation
LVIS Train	1,113	LVIS v1.0 (COCO train images)
Objects365 Val	933	Objects365 v2 validation
Total	2,470

Each annotation has exactly one human-selected 3D bounding box, chosen from candidates generated by multiple 3D estimation algorithms (LA3D, SAM3D, Algorithm, DetAny3D, 3D-MooD) and validated through a multi-stage pipeline of crowdsourced annotation, quality control, human rejection review, and geometric filtering.

Repository Structure

.
├── data/           # WildDet3D-Bench ground truth (for benchmark visualization)
│   ├── index.json          # Master index with image metadata and scene hierarchy
│   ├── boxes/              # Per-image JSON: 2D/3D boxes, categories, quality flags
│   ├── images/             # Super-resolution images (4× upscaled)
│   ├── images_annotated/   # Thumbnails with pre-rendered 3D box overlays
│   ├── camera/             # Camera intrinsic parameters
│   └── pointclouds/        # PLY point clouds (~250k points each)
│
└── model/          # Model predictions on WildDet3D-Bench (for model comparison visualization)
    ├── images/             # Images with model prediction overlays
    ├── box/                # Per-image model prediction boxes
    └── text/               # Per-image model prediction metadata

data/ — Benchmark Ground Truth

Contains the full WildDet3D-Bench validation set with human-annotated 3D bounding boxes:

• 2,196 images with at least one valid 3D annotation (274 images filtered out)
• Per-image box data includes: 2D boxes (in 4× SR coordinates), 3D boxes (10D: center + dimensions + quaternion), category names, ignore3D flags, human quality ratings
• Point clouds reconstructed from monocular depth estimation
• Annotated thumbnails with 3D boxes projected onto images, colored by object category

model/ — Model Predictions

Contains predictions from different 3D detection models evaluated on the benchmark, used by a separate model comparison visualization server.

3D Box Format

Each 3D bounding box is represented as a 10-element array:

[cx, cy, cz, w, h, l, qw, qx, qy, qz]

Field	Description
cx, cy, cz	Box center in camera coordinates (meters)
w, h, l	Box dimensions (meters)
qw, qx, qy, qz	Rotation as unit quaternion

Coordinate system: OpenCV camera convention (X-right, Y-down, Z-forward).

Annotation Pipeline

1. Monocular depth estimation — per-pixel depth maps
2. 4× super-resolution — higher quality point clouds
3. Multi-algorithm 3D box generation — candidate boxes per 2D detection
4. VLM scoring — automated quality scoring (6 criteria, 0–12 total)
5. Human annotation (Prolific) — workers select best candidate and rate quality
6. Human rejection review — second-pass review of selected boxes
7. Geometric filtering — GPT-estimated size validation and depth ratio checks
8. Composite image removal — filter collage/grid images