Recent advancements in unsupervised monocular depth estimation typically rely on an assumption that image photometry remains consistent across consecutive frames. However, this assumption often fails in endoscopic scenes due to: 1) local photometric inconsistency caused by specular reflections creating highlights; and 2) global photometric inconsistency resulting from the simultaneous movement of the light source and the camera. Since unsupervised depth estimation methods rely on appearance discrepancies between frames as a supervisory signal, these photometric inconsistencies inevitably deteriorate loss function calculation. In this paper, our goal is to obtain a strong and reliable supervisory signal for achieving photometric-consistent depth estimation. To this end, for local photometric inconsistency, we utilize the specular reflection model to introduce a Highlight Loss for handling the estimation of highlight regions. For global photometric inconsistency, we design a Photometric Match module, which utilizes the spotlight illumination model to derive an analytical expression, achieving photometric alignment across different frames. Unlike previous works that introduce additional optical flow or networks, our method is simpler and more efficient. Extensive experiments demonstrate our method achieves the state-of-the-art results on C3VD, SCARED and SERV-CT datasets.

Recently, dynamic convolution shows performance boost for the CNN-related networks in medical image segmentation. The core idea is to replace static convolutional kernel with a linear combination of multiple convolutional kernels, conditioned on input-dependent attention function. However, the existing dynamic convolution design suffers from two limitations: i) The convolutional kernels are weighted by enforcing a single-dimensional attention function upon the input maps, overlooking the synergy in multi-dimensional information. This results in sub-optimal computations of convolution kernels. ii) The linear kernel aggregation is inefficient, restricting the model's capacity to learn more intricate patterns. In this paper, we rethink the dynamic convolution design to address these limitations and propose multi-dimensional aggregation dynamic convolution (MAGIC). Specifically, our MAGIC introduce a dimensional-reciprocal fusion module to capture correlations among input maps across the spatial, channel, and global dimensions simultaneously for computing convolutional kernels. Furthermore, we design kernel recalculation module, which enhances the efficiency of aggregation through learning the interaction between kernels. As a drop-in replacement for regular convolution, our MAGIC can be flexibly integrated into prevalent pure CNN or hybrid CNN-Transformer backbones. The extensive experiments on four benchmarks demonstrate that our MAGIC outperforms regular convolution and existing dynamic convolution. Code is available at: https://github.com/Segment82/MAGIC