Activity recognition in scientific experimentation using multimodal visual encoding

Abstract

Capturing actions during scientific experimentation is a cornerstone of reproducibility and collaborative research. While large multimodal models hold promise for automatic action (or activity) recognition, their ability to provide real-time captioning of scientific actions remains to be explored. Leveraging multimodal egocentric videos and model finetuning for chemical experimentation, we study the action recognition performance of Vision Transformer (ViT) encoders coupled either to a multi-label classification head or a pretrained language model, as well as that of two state-of-the-art vision-language models, Video-LLaVA and X-CLIP. Highest fidelity was achieved for models coupled with trained classification heads or a fine-tuned language model decoder, for which individual actions were recognized with F1 scores between 0.29-0.57 and action sequences were transcribed at normalized Levenshtein ratios of 0.59-0.75, while inference efficiency was highest for models based on ViT encoders coupled to classifiers, yielding a 3-fold relative inference speed-up on GPU over language-assisted models. While models comprising generative language components were penalized in terms of inference time, we demonstrate that augmenting egocentric videos with gaze information increases the F1 score (0.52 → 0.61) and Levenshtein ratio (0.63 → 0.72, p = 0.047) for the language-assisted ViT encoder. Based on our evaluation of preferred model configurations, we propose the use of multimodal models for near real-time action recognition in scientific experimentation as viable approach for automatic documentation of laboratory work.