Optimized Small Object Detection with Large Models in Real-World Practice

This paper investigates the application of model capabilities to empower production in industrial environments through the establishment of a two-stage working mechanism. The first stage is the construction of model capabilities; the second stage is the automated deployment of data collection, data annotation, and data audit work, realizing the DFMA (Data, Model, Filter, Audit) process to achieve engineering production implementation. In the first stage, we attempted to apply traditional models such as Damo-yolo-s and Deepanything models to the detection of small targets like distant mobile phones. The target detection effects of traditional models are excellent in terms of various indicators on test datasets with similar scenarios, but they fail to recognize distant mobile phones in production scenarios with significantly different video quality or backgrounds. Based on this, we proposed the use of large models to enhance the detection capability of small targets. Our experiments show that large models fine-tuned with LoRA achieve a generalization recognition accuracy of 100%, a precision of 97.8%, and a recall rate of 93% in actual production landing scenarios, effectively solving the problem of difficult small target recognition. In the second stage, we established the DFMA workflow to address the challenges of data annotation and annotation data audit in industrial landing scenarios, and this process has been implemented in the production environment.