Eş zamanlı çıkarım için ikili ağların monoküler derinlik tahminine uyarlanması

Monocular Depth Estimation (MDE) is a fundamental computer vision application area for many industry-related advances. Due to its deployment needs, the inference time of the depth estimation algorithm also plays a crucial role among other accuracy metrics. With the recent advances in Convolutional Neural Networks (CNNs) on other time-constrained computer vision tasks, many efficient feature extractors have been studied and adopted from MDE models as the backbone. Although those feature extractors have shown significant improvement in throughput, the widely-used encoder-decoder architecture used by Real-time MDE models also relies on a decoder network for upsampling. Following a similar approach, stacking multi-channel convolutional layers on a decoder hinders the inference time. This study investigates the benefits of Bilateral Networks in Real-time MDE tasks. During our research, we first manipulate the structure of a recently introduced real-time segmentation model (STDC-Seg) for the MDE problem. Once we attain real-time inference speed, we tailor the backbone structure and attention modules of the model for the needs of MDE to improve prediction accuracy. Finally, we train the models on the well-known KITTI dataset and compare our results with the models of the KITTI Eigen Split MDE Benchmark along with the previous real-time models. Our experimental results show that our real-time method achieves on-par metric performance with state-of-the-art models that are not subject to any time-constraint.