Exploring Real-Time Model Augmentation and Pretraining for Optimized Road Damage Detection

Road safety remains a significant global concern, and computer vision techniques are increasingly being deployed for automated damage assessment and hazard mitigation. However, road surface damage detection is particularly challenging to generalize, especially in resource-constrained environments. In this work, we approach road damage detection using RTMDet, a one-stage real-time object detection model. RTMDet offers an optimal balance of inference speed, memory efficiency, and accuracy for industrial detection tasks. The RTMDet model evaluated 5% higher mAP50 than the popular YoloV8 model on CRDDC 2022 data and demonstrated robust learning over longer augmented training cycles. Our evaluation introduces several strategies to enhance detection performance: (a) We propose the CSPNeXt backbone as a superior alternative to the Swin Transformer for damage detection; (b) We implement an augmentation pipeline stage-switching strategy, transitioning from complex multiscale transforms to simpler affine transforms in the final training epochs, which accelerates model convergence; (c) We demonstrate the effectiveness of transfer learning by pretraining on the CRDDC 2022, external datasets, and improving generalization in large model; (d) On a resolution-specific model with all performance-enhancing strategies, we observe 14% improvement in mAP50 on low-resolution images showcasing the challenge high-resolution images present in this dataset. Our best-performing models include RTMDet-Large pretrained with a classification task and RTMDet-Medium detector models, strategically trained on competition data only. The Phase 1 submission leaderboard test evaluation for RTMDet-Large is 0.57 and RTMDet-Medium is 0.58 F1 score. The Phase 2 inference speed test evaluation on all countries’ test images for large and medium models are 0.0704 and 0.0491 seconds/image respectively. The inference speed evaluation is done in a containerized model deployed on a standard AWS EC2 GPU (g4dn.xlarge) server environment. All source code, instructions, and model parameters are publicly released at https://github.com/karmarv/orddc2024.