CV Data Sets

2025

• ETH3D
• Aria Gen 1/2
• 📍 Kubric - synthetic
• Nature - Fair human-centric image dataset for ethical AI benchmarking
• Human Mesh Modeling for Anny Body - NAVER LABS Europe
• 2025 - Kinematify: Open-Vocabulary Synthesis of High-DoF Articulated Objects
• 2025 - Automatic analysis of three-dimensional cardiac tagged magnetic resonance images using neural networks trained on synthetic data
• SwissHeart Study

Robots

• 2021 - Habitat - Matterport 3D Research Dataset

2026

4D Reconstruction

Dataset Specifications

Dataset	Data Type	Depth Map	Semantic / Mask	Camera Pose	Temporal Length	Approx Size
PointOdyssey v1.2 (Stanford, ICCV 2023)	Synthetic, long-term deformable human/animal scenes, 30 fps, 540×960	Dense depth maps and surface normals per frame	Instance segmentation and visibility mask	Full camera intrinsics and extrinsics	~2 000 frames per video × 159 videos (≈ 1 h 30 min)	≈ 300 GB
Dynamic Replica v2 (Meta AI, CVPR 2023 DynamicStereo)	Synthetic RGB-D videos of human–object interactions	Dense depth maps	Semantic segmentation masks	Full camera poses	~300 frames (≈ 10 s) per video × 524 videos (≈ 1 h 15 min training split)	≈ 1.7 – 1.8 TB
Kubric (Google Research 2022)	Procedural synthetic multi-object MVS generator	Per-pixel depth and normals	Segmentation and instance masks	Intrinsics and extrinsics available	Short clips (≈ 24 frames average)	≈ 1 TB
MPI-Sintel (Complete)	Blender-rendered movie frames with optical-flow ground truth	Depth and optical flow maps	Semantic layers (albedo, shading)	Camera poses provided	23 training + 12 test sequences (≈ 1 min total)	≈ 5.3 GB
TUM-Dynamics (RGB-D SLAM Benchmark 2012)	Real RGB-D video sequences with moving objects	Depth maps from Kinect sensor	No semantic mask provided	Ground-truth camera poses	2 – 3 min per sequence × 15 – 20 sequences (≈ 1 h total)	≈ 60 GB
ETH3D (Schöps et al., CVPR 2017)	Real multi-view stereo benchmark (static scenes)	Ground-truth depth maps	No semantic mask provided	Calibrated multi-view poses	Static scenes (~10 – 50 images per scene)	≈ 25 GB

NeurIPS 1987–2024

Period	Core Theme	Representative Focus
1987–1993	Neuro-inspired Foundations	Neural computation, Hebbian learning, Hopfield networks
1994–2000	Statistical Learning Era	SVM, VC theory, kernel methods
2001–2009	Probabilistic Modeling	Graphical models, Bayesian inference, LDA, GP
2010–2014	Representation Revival	Deep belief nets, optimization, sparse coding
2015–2017	Theoretical Consolidation	Learning theory, optimization, multi-agent RL
2018–2019	Deep Learning Formalization	Neural ODEs, distributed optimization, RL theory
2020–2021	Scaling Reflection	GPT-3, robustness, evaluation metrics
2022–2023	Diffusion + LLM Revolution	Generative diffusion, alignment, safety
2024→	Multimodal + Reasoning Integration	Vision-language models, efficiency, real-world benchmarks