Utilizing Stable Diffusion to Enhance Car Parts Detection

Car parts detection necessitates a substantial amount of up-to-date data and requires detectors with powerful feature learning capabilities. To this end, we propose two novel methods that leverage the high-level features learned from Stable Diffusion for data augmentation and car parts detection. Specifically, we introduce Stable Diffusion Augmentation (SDA) and Inpainting Stable Diffusion Augmentation (ISDA), which aim to generate or modify a broader range of innovative car part samples. Additionally, we present the Diffusion-Based Detector (DBD), which utilizes latent diffusion to extract multi-level representations. These representations are subsequently fed to the detector head for localization and classification tasks. Experimental results demonstrate that SDA and ISDA effectively enhance the semantic information of specific car parts, while the augmented data significantly improves the performance of car parts detection, with a 3% increase compared to basic augmentation. Moreover, DBD achieves higher Mean Average Precision (mAP) in car parts detection, 6% higher than classic detectors. Overall, the experimental outcomes validate the efficacy of our proposed method in the domain of car parts detection.