Learning-based Multi-View Stereo (MVS) methods heavily rely on feature extraction and cost volume construction to bridge 2D semantics and 3D spatial associations. However, many recent studies have increasingly shifted focus away from these two critical steps, often adopting a multi-stage cascaded framework that can propagate errors from earlier stages. To this end, we introduce TTINS-MVSNet, a novel progressive refinement framework that integrates a one-stage MVS module enhanced by two types of Epipolar Transformer (ET) and subsequent depth optimization modules. Specifically, we incorporate an intra-view ET into context-aware feature extraction and an inter-view ET into visibility-aware cost aggregation. A significant innovation in our approach is the introduction of implicit neural super-resolution module, designed to recover finer details. Experimental results on benchmark datasets demonstrate that our method outperforms all previous approaches with similar structures, highlighting its effectiveness and generalization capability. The code will be available after publication.

We propose an instance segmentation model, named CoMask, that effectively alleviates the scale variation issue and addresses the precise localization. Specifically, we develop a multi-scale feature extraction module (MSFEM) to exploit multi-scale spatial cues. Besides, the channel attention mechanism is also adopted to further enhance the discriminating ability. Equipped with MSFEMs, multi-scale and multi-level features can be extracted to better characterize objects of various sizes and provide affluent high-level semantic information. For precise localization, we propose a collaborative learning framework to compute coarse masks and regresses position-sensitive dense offsets. The foreground confidence of each position is then assigned as the weight of the corresponding bounding box to calculate a weighted average. Thus, we can mitigate interference of background regions. After obtaining the final regressed bounding boxes, finer foreground masks can be calculated. We conduct experiments on MS COCO dataset. Experimental results validate that CoMask is competitive compared with state-of-the-art models.

Research on designing local feature descriptors has gradually shifted to deep learning. Different from other computer vision tasks, the biggest challenge for local descriptor learning lies with the formulation of loss functions. Existing methods solve the problem by leveraging Siamese loss or triplet loss and improve the performance of the learned descriptors by a significant margin. However, the widely used Siamese loss and triplet loss cannot fully utilize the context information. In this paper, we propose a novel loss function to introduce more context information to facilitate training. After incorporating the proposed loss function into training, our learned descriptor demonstrates state-of-the-art performance in patch verification, image matching and patch retrieval benchmarks. The pretrained model will be publicly available at https://github.com/clelouch/CANet.