Customizable Computer Vision Expands Data Access Without Compromising Privacy
summary
Open-source Perception package and People Sans People provide a CLI, Mac/Linux binaries, rich labels (boxes, 3D boxes, seg, keypoints, depth, lidar) and programmatic scene control. Pre-training on 490,000 synthetic images then fine-tuning on COCO yields higher bounding-box and 17‑keypoint accuracy than ImageNet pre-training, with the biggest gains when only 5k–10k real images are available. Zero‑shot performance remains weak but synthetic pretraining improves out‑of‑distribution generalization across multiple human‑centric datasets.
AI-written summary · about
chapters (9)
- 0:00Synthetic data in computer vision
- 1:50Training with real-world data
- 3:14Domain randomization for synthetic data
- 6:05Requirements for realistic synthetic content
- 7:08Perception package and human-centric tools
- 8:54Controlling synthetic data distributions
- 12:32Evaluating synthetic data
- 15:39Generalization of synthetic models
- 17:24Open-source synthetic data tools
related talks
23mBuilding high-quality computer vision models using only a few examples
58mUsing Unity to Generate Synthetic data and Accelerate Computer Vision Training
01h 29mHuman-like Visual Search Application with Small Data
01h 18mComputer Vision after the Victory of Data
41mKEYNOTE: Image processing, AI, and autonomous systems
01h 07mComputer Vision with PyTorch
40mThe Effortless Development of Custom Computer Vision Models
01h 00mOpening Keynote