Research Assistant - Computer Vision Eckerlab
I research video generalization and object bias in action-recognition models. I benchmarked the Something-Something-v2 dataset against a custom multi-camera distractor dataset and engineered a synchronized three-camera hardware setup to isolate spatial object bias from temporal motion.
I built multimodal feature extraction and cross-attention probing pipelines for frozen vision foundation models, including DINOv2 and V-JEPA2. I also developed a SAM3-based automated object-occlusion workflow for approximately 150,000 frames and evaluated a 320B-parameter vision-language model through cloud APIs using structured JSON prompts and base64 video representations.
I trained two-stream I3D video classification architectures using RGB and optical-flow inputs for cross-domain generalization. I created modular PyTorch training workflows with Weights & Biases tracking, early stopping, learning-rate scheduling, checkpointing, HDF5 data serialization, and late-fusion inference, improving uAP from 72.8 to 75.2 on a heavily imbalanced dataset.