Unleashing the Power of Motion and Depth: A Selective Fusion Strategy for RGB-D Video Salient Object Detection

Applying salient object detection (SOD) to RGB-D videos is an emerging task called RGB-D VSOD and has recently gained increasing interest, due to considerable performance gains of incorporating motion and depth and that RGB-D videos can be easily captured now in daily life. Existing RGB-D VSOD models have different attempts to derive motion cues, in which extracting motion information explicitly from optical flow appears to be a more effective and promising alternative. Despite this, there remains a key issue that how to effectively utilize optical flow and depth to assist the RGB modality in SOD. Previous methods always treat optical flow and depth equally with respect to model designs, without explicitly considering their unequal contributions in individual scenarios, limiting the potential of motion and depth. To address this issue and unleash the power of motion and depth, we propose a novel selective cross-modal fusion framework (SMFNet) for RGB-D VSOD, incorporating a pixel-level selective fusion strategy (PSF) that achieves optimal fusion of optical flow and depth based on their actual contributions. Besides, we propose a multi-dimensional selective attention module (MSAM) to integrate the fused features derived from PSF with the remaining RGB modality at multiple dimensions, effectively enhancing feature representation to generate refined features. We conduct comprehensive evaluation of SMFNet against 19 state-of-the-art models on both RDVS and DVisal datasets, making the evaluation the most comprehensive RGB-D VSOD benchmark up to date, and it also demonstrates the superiority of SMFNet over other models. Meanwhile, evaluation on five video benchmark datasets incorporating synthetic depth validates the efficacy of SMFNet as well. Our code and benchmark results are made publicly available at https://github.com/Jia-hao999/SMFNet.

R-Stitch: Dynamic Trajectory Stitching for Efficient Reasoning

Chain-of-thought (CoT) reasoning enhances the problem-solving capabilities of large language models by encouraging step-by-step intermediate reasoning during inference. While effective, CoT introduces substantial computational overhead due to its reliance on autoregressive decoding over long token sequences. Existing acceleration strategies either reduce sequence length through early stopping or compressive reward designs, or improve decoding speed via speculative decoding with smaller models. However, speculative decoding suffers from limited speedup when the agreement between small and large models is low, and fails to exploit the potential advantages of small models in producing concise intermediate reasoning. In this paper, we present R-Stitch, a token-level, confidence-based hybrid decoding framework that accelerates CoT inference by switching between a small language model (SLM) and a large language model (LLM) along the reasoning trajectory. R-Stitch uses the SLM to generate tokens by default and delegates to the LLM only when the SLM's confidence falls below a threshold. This design avoids full-sequence rollback and selectively invokes the LLM on uncertain steps, preserving both efficiency and answer quality. R-Stitch is model-agnostic, training-free, and compatible with standard decoding pipelines. Experiments on math reasoning benchmarks demonstrate that R-Stitch achieves up to 85\% reduction in inference latency with negligible accuracy drop, highlighting its practical effectiveness in accelerating CoT reasoning.

FPSAttention: Training-Aware FP8 and Sparsity Co-Design for Fast Video Diffusion

Diffusion generative models have become the standard for producing high-quality, coherent video content, yet their slow inference speeds and high computational demands hinder practical deployment. Although both quantization and sparsity can independently accelerate inference while maintaining generation quality, naively combining these techniques in existing training-free approaches leads to significant performance degradation due to the lack of joint optimization. We introduce FPSAttention, a novel training-aware co-design of FP8 quantization and sparsity for video generation, with a focus on the 3D bi-directional attention mechanism. Our approach features three key innovations: 1) A unified 3D tile-wise granularity that simultaneously supports both quantization and sparsity; 2) A denoising step-aware strategy that adapts to the noise schedule, addressing the strong correlation between quantization/sparsity errors and denoising steps; 3) A native, hardware-friendly kernel that leverages FlashAttention and is implemented with optimized Hopper architecture features for highly efficient execution. Trained on Wan2.1's 1.3B and 14B models and evaluated on the VBench benchmark, FPSAttention achieves a 7.09x kernel speedup for attention operations and a 4.96x end-to-end speedup for video generation compared to the BF16 baseline at 720p resolution-without sacrificing generation quality.

Salient Object Detection in RGB-D Videos

Given the widespread adoption of depth-sensing acquisition devices, RGB-D videos and related data/media have gained considerable traction in various aspects of daily life. Consequently, conducting salient object detection (SOD) in RGB-D videos presents a highly promising and evolving avenue. Despite the potential of this area, SOD in RGB-D videos remains somewhat under-explored, with RGB-D SOD and video SOD (VSOD) traditionally studied in isolation. To explore this emerging field, this paper makes two primary contributions: the dataset and the model. On one front, we construct the RDVS dataset, a new RGB-D VSOD dataset with realistic depth and characterized by its diversity of scenes and rigorous frame-by-frame annotations. We validate the dataset through comprehensive attribute and object-oriented analyses, and provide training and testing splits. Moreover, we introduce DCTNet+, a three-stream network tailored for RGB-D VSOD, with an emphasis on RGB modality and treats depth and optical flow as auxiliary modalities. In pursuit of effective feature enhancement, refinement, and fusion for precise final prediction, we propose two modules: the multi-modal attention module (MAM) and the refinement fusion module (RFM). To enhance interaction and fusion within RFM, we design a universal interaction module (UIM) and then integrate holistic multi-modal attentive paths (HMAPs) for refining multi-modal low-level features before reaching RFMs. Comprehensive experiments, conducted on pseudo RGB-D video datasets alongside our RDVS, highlight the superiority of DCTNet+ over 17 VSOD models and 14 RGB-D SOD models. Ablation experiments were performed on both pseudo and realistic RGB-D video datasets to demonstrate the advantages of individual modules as well as the necessity of introducing realistic depth. Our code together with RDVS dataset will be available at https://github.com/kerenfu/RDVS/.