@MASTERSTHESIS{ 2021:672064314,
 	title = {Convolutional neural networks compression for object detection},
 	year = {2021},
 	url = "http://tede2.pucrs.br/tede2/handle/tede/9890",
 	abstract = "Deep Learning (DL) is the state-of-the-art in Computer Vision tasks, such as Image Classification, Object Detection, Instance Segmentation, Content Generation, among others. Over time, the models have become broader, deeper, and more accurate, but also hyperparameterized, heavier, and slower, making their use harder for automating tasks based on constrained devices, such as those with reduced processing power, or with memory or energy consumption constraints. Consequently, Model Compression emerges in the literature to reduce the model’s size and processing cost as much as possible, while impacting as little as possible in the model’s performance within its target task. Although there are many model compression studies in the literature exploring several different  approaches, there are few studies in the literature bringing practical comparisons between different approaches and none of those focusing on Object Detection. Therefore, this work contributes to the literature by comparing and exploring the existing trade-offs between Pruning, Knowledge Distillation (KD), Neural Architecture Search (NAS), and a model reconstruction based on efficient convolutions. To achieve this goal, we train models based on YOLOv3 with the same data augmentation on two datasets, PASCAL VOC and Exclusively Dark Images, and we evaluate them according to Mean Average Precision, number of parameters, storage size, and Multiply-Accumulate  operations (MACs). Results show that a more aggressive Pruning was capable of generating the best trade-off: its mAP surpassed a NAS + KD approach, in addition to producing a model with the smallest number of parameters and with a most effective reduction in MACs.",
 	publisher = {Pontifícia Universidade Católica do Rio Grande do Sul},
 	scholl = {Programa de Pós-Graduação em Ciência da Computação},
 	note = {Escola Politécnica}
}