Wavelet Packing for Self-Supervised Monocular Depth Estimation

We propose a novel self-supervised convolutional neural network for estimating pixel-wise depth from monocular camera images. To preserve detailed information while expanding receptive fields within an encoder-decoder architecture, we utilize the wavelet transform for information packing (Wav-Packing) in the encoder and the inverse wavelet transform for information unpacking (WavUnpacking) in the decoder. The proposed depth estimation network, termed WavPackNet, is integrated into a self-supervised structure-from-motion learning framework, enabling training on unlabeled image sequences. Comprehensive experiments conducted on the CityScapes and KITTI datasets demonstrate that WavPack-Net achieves superior performance compared to 3D PackNet in most configurations, while significantly reducing both training and inference time. Code and pre-trained models are released at https://github.com/leiqin1/WavPackNet.