Our work aims to bridge the gap between images of handwriting and digital ink with a Vision Language Model (PaLI). To our knowledge, this is the first work that effectively does so with arbitrary photos with diverse visual characteristics and backgrounds. Furthermore, it generalizes beyond its training domain and can work on simple sketches. Human evaluation reveals that 87% of the samples produced by our model on the challenging HierText dataset are considered valid tracings of the input image, and 67% look like pen trajectories traced by a human.

We introduce Modular Quantization-Aware Training (MQAT), an adaptive and mixed-precision quantization-aware training strategy that exploits the modular structure of modern 6D pose estimation architectures. MQAT guides a systematic gradated modular quantization sequence and determines module-specific bit precisions, leading to quantized models that outperform those produced by state-of-the-art uniform and mixed-precision quantization techniques.

This project investigates the effects of Sharpness-Aware Minimization (SAM) and Adaptive Sharpness-Aware Minimization (ASAM) on model generalization. Our experiments demonstrate that sharpness-aware optimization techniques notably enhance generalization abilities. Notably, ASAM shows promise in improving performance on un-normalized data.