Enhancing Shape Bias in CNNs
Combining depth density estimation with image classification to enhance shape bias in CNNs.
Project Overview
LOCNet explores enhancing shape bias in CNNs by jointly optimizing depth estimation and classification tasks. Using a ResNet50-based encoder-decoder architecture trained on a novel ImageNet-derived dataset, our model outperforms classification-only models on out-of-distribution datasets, particularly in sketch and texture-based filters. This approach shows promise in bridging the gap between human and machine vision, improving robustness in image classification tasks.