A Hardware-Aware Sampling Parameter Search for Efficient Probabilistic Object Detection

Recent advancements in Deep Neural Networks (DNNs) have led to remarkable achievements in object detection, making them increasingly relevant for safety-critical domains like autonomous driving. However, a significant challenge for deploying DNN-based object detection in safety-critical applications remains the inability of the models to estimate their own uncertainty. To address this issue, Probabilistic Object Detection has emerged as a solution, allowing for the assessment of both semantic and spatial uncertainty. Monte-Carlo sampling methods, such as Dropout and its variants, are commonly used to generate the necessary probability distributions. Nonetheless, determining the appropriate Dropout variant, sample size, and drop probability for a specific probabilistic model remains a complex task, especially when considering the importance of balancing algorithmic accuracy and hardware efficiency. Ensuring hardware efficiency is particularly crucial for deploying these models in embedded systems. To tackle this challenge, we treat it as an optimization problem and employ an evolutionary multi-objective search to identify the best-fitting sampling parameters. In our evaluation using the YOLOv5 model, we demonstrate that Gaussian Dropout outperforms other Dropout variants. Notably, we achieve a doubling of the PDQ score with no retraining and an mAP50- 95 loss of only 1% on the COCO dataset. Additionally, our study unveils the non-intuitive trade-offs considering hardware performance.