Using Synthetic Data For Object Detection on the edge in Hazardous Environments

Abstract: This thesis aims to evaluate which aspects are important when generating synthetic data with the purpose of running on a lightweight object detection model on an edge device. The task we constructed was to detect Canisters and whether they feature a protective valve called a Cap or not (called a No-Cap). The problem was split into three separate classes in order to evaluate objects with varying size and complexity. Canister was the largest class while Cap and Nocap where smaller classes where No-cap is of a higher degree of complexity. The choice of model focused on both performance and inference speed, thus we chose SSD MobilNet V2 which is pre-trained on the Coco dataset. Additionally The model is quantized in order to perform better on edge devices. To test important features and techniques for generating synthetic data we conducted three experiments. First we test the importance of a scene structure by creating a hyperrealistic dataset with scene structure and one without. The results indicate that scene structure is important and can encode patterns which aids the model in object detection. Then we tested the importance of Domain Randomization where the material of different assets of the canisters were randomized. The dataset with randomized material of the canister body produced the best results. It seems that the model extracts features of the smaller assets more efficiently if the background object is randomized. The third test analyzed the importance of data augmentation and having the model pre-trained. A pre-trained model was not shown to be necessary and training from scratch proved to be more advantageous the more randomized the dataset is. It was also noted that data augmentation was crucial for object detection when it comes to synthetic data. Without augmentation the model failed to detect anything but the canister body. This was the case for both hyperrealistic and domain randomized datasets. The best results were achieved by training the model from scratch using the dataset of randomized canister body material and using augmentation.