Roadside animal detection

In this work, I trained SSD based object detection algorithms to detect animals on road and explored what impact do the different feature extractor architecture, artificially generated animal images and different training tweaks make on a SSD based object detection algorithms accuracy and inference speed. The changes in MobileNetV2 feature extractors convolutional layers widths did not result in optimal accuracy and inference speed trade-off. Widening the layers caused the model to make slightly more accurate animal predictions, but at a lower pace. After reducing the layers widths, the model gained considerable inference speed boost both on desktop GPU and mobile CPU, but at the cost of substantial amount of accuracy. The animal detection on road problem requires the model to make as accurate predictions as possible at the highest rate possible, so sacrificing a lot of inference speed for minimal accuracy improvement, or vice versa, was not the optimal solution. Due to lack of specific animal images to train models on, I decided to write an algorithm to artificially generate those images. After adding 100 artificially generated images to my training dataset, the model demonstrated 0.87% improvement in accuracy. In order to further improve animal detection accuracy, I implemented image mixup data augmentation to use while training the models. By using image mixup with beta distribution B(1, 1) the model reached 8.02% accuracy improvement. Additionally, I used brightness, contrast, hue, saturation and horizontal image flip image augmentations, that were randomly applied to images while training the models. Only the hue augmentation did not have measurable positive impact on model’s prediction accuracy when using it as a stand-alone training data augmentation. But when using all these augmentations, including hue, the model predicted animals with more accuracy, than the model which did not use hue augmentation. Lastly, when using all the measures that improved the animal detection accuracy, namely generated animal images, image mixup and all the augmentations, the model demonstrated 9.46% increase in accuracy