Omnidirectional videos impose intensive demands on efficient coding due to high spatial resolution and unique projection characteristics. This paper investigates spherical rotation angle prediction for omnidirectional video coding and proposes a content-adaptive spherical rotation algorithm that fuses motion and texture characteristics under the joint optimization framework of Equi-Angular Cubemap (EAC) and spherical rotation. Specifically, an adaptive motion-intensity classification scheme is designed to categorize videos into high- and low-motion classes based on average motion energy. For low-motion videos, texture-gradient analysis is utilized to achieve robust and low-complexity rotation angle prediction. For high-motion videos, a motion vector (MV) statistics-based method with a magnitude-weighted directional histogram is designed to accurately estimate the dominant rotation angle. The proposed dual-branch method avoids pre-encoding and maintains low computational complexity. Experimental results demonstrate that the proposed scheme achieves an average bitrate saving of 0.52% and a maximum bitrate saving of 1.89% for specific test sequences compared to EAC. This work validates the effectiveness of joint projection-rotation optimization and provides a practical solution for efficient omnidirectional video coding.
Learned image compression leverages neural networks to transform input images into compact latent representations for improved compression performance. Most existing methods employ a fixed latent spatial scale across all inputs. However, natural images exhibit diverse spatial information densities, making fixed-scale representations potentially suboptimal. To address this limitation, we propose a content-adaptive spatial scale routing (CASR) mechanism that dynamically selects the optimal latent spatial scale for each image within a unified model. Specifically, we introduce switchable scale branches through scale-specific projection and reconstruction operators while sharing a common transform backbone across different branches. During encoding, the optimal branch is selected by minimizing RD cost. The coexistence of differentiable optimization and discrete scale selection makes end-to-end training challenging. To address this issue, we further propose a joint training strategy that alternates between routing assignment optimization and gradient-based network parameter updates. Furthermore, to reduce routing complexity during encoding, we propose a fast routing strategy to predict the optimal branch using only a single forward pass. Experimental results show that CASR achieves superior compression performance while substantially reducing decoding complexity compared with fixed-scale baselines. On DIV8K, CASR achieves over 4% bitrate savings and over 30% decoding complexity reduction.
Weakly supervised camouflaged object detection (WSCOD) remains challenging because scribble annotations provide limited object coverage for supervision. Although recent methods leverage the Segment Anything Model (SAM) with point or scribble prompts to generate pseudo-labels, the resulting pseudo-labels are biased toward interior regions, missing boundary details and low-contrast object parts. To address these problems, we propose TACO, a Text and Point Collaborative Pseudo-Labeling framework with Boundary-Semantic Guidance for WSCOD. To obtain reliable pseudo-labels, we propose a Text-Point Collaborative Pseudo-Labeling strategy (TPCP). Specifically, we introduce Adaptive Point Optimization to improve scale-aware localization from point prompts, and use textual descriptions as complementary semantic cues to recover regions missed by sparse point coverage. While the generated pseudo-labels become more complete, boundaries and ambiguous regions remain challenging in weak supervision. To address this issue, we further introduce a Boundary-Semantic Guided Feature Learning network, where a Boundary-Guided Reasoning Module (BGRM) refines object boundaries by modeling region-boundary interactions, and an Adaptive Text-Guided Fusion Module (ATFM) uses textual semantics to complement visual features with high-level target semantics. Extensive experiments on CAMO, COD10K, and NC4K demonstrate that the proposed method outperforms existing state-of-the-art weakly supervised methods.
While RGB-Thermal (RGBT) tracking theoretically promises robust all-weather observation, existing approaches often fail to establish stable semantic anchors, rely on lossy spatial alignments that disrupt representational manifolds, and lack flexible aggregation to prevent feature contamination. To address these limitations, we propose DYTrack, a progressive injection framework for RGBT Tracking. Specifically, we introduce a Semantic-Locked Query Injection (SLQI) mechanism that leverages pre-trained linguistic priors as stable anchors. This extracts modality-invariant cues to effectively mitigate initial visual degradation. To establish rigorous spatial correspondence, we design a Y-shaped Spatial Identity Injection (Y-SII) module that enforces gradient stability to preserve the spatial topology. This yields a pristine, pooling-free alignment derived strictly from token-to-token interactions. By capitalizing on this aligned foundation, we present a Dual-Gate Deformable Aggregation (DDA) strategy that dynamically sculpts continuous grids via synergistic regularizations, robustly safeguarding the aggregation from spurious artifacts. Extensive experiments on LasHeR, RGBT234, GTOT, and VTUAV benchmarks demonstrate that our DYTrack consistently outperforms state-of-the-art methods while maintaining real-time speed.
Snapshot compressive hyperspectral reconstruction aims to recover three-dimensional hyperspectral images from two-dimensional compressed measurements formed by optical dispersion and masking. Recent Transformer-based methods have achieved promising reconstruction performance by modeling long-range dependencies with self-attention, but their global feature interactions incur substantial computational and memory costs. Shift operations provide an efficient alternative for receptive-field expansion through lightweight feature transport. However, existing shift designs usually move features along generic spatial directions, whereas CASSI dispersion induces wavelength-dependent displacements only along a fixed direction. This mismatch may lead to redundant feature interactions and weaken the aggregation of useful dispersion-related context. To address this issue, we propose ShiftSCI, a fully convolutional unfolding network that incorporates dispersion-guided shifting for efficient hyperspectral image reconstruction. At the core of ShiftSCI is a dual-branch Dispersion-Guided Shift Module (DGSM), where the shifted branch aggregates dispersion-related context and the unshifted branch preserves local structures to reduce shift-induced artifacts. Beyond dispersion-related context aggregation, accurate CASSI reconstruction also requires recovering fine spatial details and maintaining local spectral-feature consistency under severe compression. We therefore introduce a Spatial–Spectral Refinement Module (SSRM), which enhances spatial textures with multi-scale convolutions and refines spectral-feature interactions through grouped channel mixing. Extensive experiments on simulated and real CASSI datasets demonstrate that ShiftSCI achieves competitive reconstruction accuracy with lower computational complexity than the compared Transformer-based methods, providing an efficient dispersion-guided solution for hyperspectral snapshot compressive reconstruction.
