[2505.01743] An LLM-Empowered Low-Resolution Vision System for On-Device Human Behavior Understanding

[Submitted on 3 May 2025]

PDF view of the paper entitled LLM LLM LLM to understand human behavior on the device, by Siyang Jiang and 10 other authors

PDF HTML (experimental) view

a summary:The rapid developments in the LVLMS language models provide the ability to overcome traditional signs by generating more richer and more detailed descriptions to understand human behavior on the device (HBU) in low -resolution vision systems, such as depth, heat and infrared. However, the current large LVLM language model methods are unable to understand low -precision data as they are designed primarily for high -resolution data, such as RGB images. The rapid installation approach is to be an illustrative designation as a large amount of low -resolution data, but it requires a large amount of intensive educational comments efforts. In this paper, we suggest a new system to provide employment, llambda, designed to support low -resolution HBU. The basic idea is to take advantage of the limited data called and a large amount of data not called LLMS to generate media comments, which can be combined with the initial data of LVLM models effectively to understand low -resolution videos in HBU. First, we suggest a data poster directed towards the contradiction, which can capture behavior -related information from low -resolution videos and create high -quality false stickers for unprecedented data through contrast learning. Second, we suggest an explanation of guides in material knowledge, which uses spatial and temporal consistency to relieve errors in false stickers. Therefore, LLMS understanding can be improved for serial data and then creating high -quality video comments. Finally, to ensure the device’s deployment, we use an effective setting in Lora to LVLMS air conditioning of low accuracy data. We evaluate Llambda using a real test across the region and three distinct, low -resolution data sets, and experiments show that Llambda excels over many modern LVLM systems of up to $ 40.03 \ % $ in average bert.