A ShuffleNetV2 CNN model was trained to estimate 3D object sizes using 2D images.

To improve performance of the base model, the Convolutional Block Attention Module (CBAM) and image denoising/edge enhancement techniques were integrated during training. It was found that while additional features improved base performance for object height and width predictions, the model still struggled to estimate object depths accurately.

This project is primarily focused on monocular depth estimation, since only 2D images alongside bounding box annotations are used for model training. Depth estimation accuracy is likely to improve significantly if additional training annotations that provide camera to object distances are utilized (e.g. LiDAR).