Generating robust real-time object detector with uncertainty via virtual adversarial training

Despite remarkable accuracy improvement in convolutional neural networks (CNNs) based object detectors, there are still some problems in applying on some safety-critical domain, such as the self-driving domain, in part due to the complexity of verifying the correctness of detecting results and the lack of safety guarantees. By simply modeling the bounding box parameters with a Gaussian distribution in a real-time object detector, we propose a new method for predicting uncertainty, which can quantify the reliability of the neural networks' prediction, to validate the correctness of detecting results with low computational complexity. In addition, we redesign the loss function by adding a new regularization term, called virtual adversarial training (VAT). The use of VAT, which is defined as the robustness of the conditional label distribution around input data against local perturbation, can smooth the output distribution robust with lower uncertainty and the prediction from the regularized model will be better. In consideration of the trade-off between the size and speed, we choose some lightweight models as the backbone of a YOLOv3 detector and the experimental results on PASCAL VOC dataset and MS COCO demonstrate the effectiveness of the proposed approach.

Automated summary

This study introduces a new method for modeling bounding box parameters with a Gaussian distribution to quantify the reliability of neural network predictions, and redesigns the loss function by adding virtual adversarial training to improve model prediction performance. Lightweight models were chosen as the backbone of the detector in experiments, demonstrating the effectiveness of the proposed approach.