Juan Pablo Oberhauser

Computer vision scientist developing behavioral phenotyping systems for laboratory mice in preclinical research, spanning detection, tracking, pose estimation, and behavioral classification.

Interested in: self-supervised pretraining, multi-task learning, contrastive learning, masked pretraining, vision-language models, few-shot learning, temporal sequence modeling, knowledge distillation, transformer architectures, active learning.

Publications

Open courses I am creating

I built these because I couldn't find ground-up resources on these topics. Each one is a full skeleton with exercise notebooks and worked solutions.

Vision Transformers & Self-Supervised Learning
Hands-on course covering ViT architectures and modern SSL methods.
 View source on GitHub →
Vision-Language Course
Self-paced material on multimodal vision-language models.
 View source on GitHub →

Projects

pybaseball
Fork of pybaseball supercharged with LLMs for easy natural language question answering! Pull current and historical baseball statistics using Python (Statcast, Baseball Reference, FanGraphs).
github.com/jpoberhauser/pybaseball
Multi-Object Trackers Collection
Keeping up with multi-object trackers. A collection, organized by type, of the latest trackers, ReID, and surveys — plus all the building blocks of trackers.
github.com/jpoberhauser/multi-object-trackers-collection
baseballCompanion
Ask natural language questions about the current state of baseball, grounded in recent YouTube analysis videos. Uses Whisper for transcription, SentenceTransformers for embeddings, FAISS as the vector store, and llama.cpp for local LLM inference with RAG (currently Mistral 7B Instruct, swappable backends). No API keys required.
github.com/jpoberhauser/baseballCompanion

Recent notes

See all posts →