SIGMA-Lane: Scale-pyramId Gated MAmba for Temporally Consistent Video Lane Detection
Authors: Tiancheng Zhang, Mengmeng Wang, Yan Gao, Xiangjie Kong, Guojiang Shen, Jiaxin Du
Organizations: Zhejiang Key Laboratory of Visual Information Intelligent Processing, College of Computer Science and Technology, Zhejiang University of Technology
Video lane detection requires predictions that remain stable across frames, yet severe vehicle occlusions can break temporal cues. In streaming recurrent models, corrupted observations may enter the hidden state and produce errors that persist into later frames. Existing occlusion-aware refinements usually provide obstacle masks as auxiliary inputs, so the state-update path is only indirectly protected. We propose SIGMA-Lane, which treats this failure mode as state contamination in State Space Model (SSM)-based temporal modeling. SIGMA-Lane places occlusion-aware gates on the SSM write and residual-fusion paths, controlling how current observations enter temporal memory and are fused back after temporal propagation. After coordinate-consistent affine alignment, the model combines two complementary paths: SSM-consistent dual-gating for temporal filtering and Structural Spatial Retrieval (SSR) for recovering missing lane structure from aligned historical priors. Experiments on VIL-100 and OpenLane-V show improved temporal stability under heavy occlusion, with competitive F1 and mIoU scores.
Figures & tables
Figure 1: Comparison of Video Lane Detection Paradigms. The window-batched, joint-aggregation paradigm (a) extracts features Ft−T+1:t from a fixed temporal window and aggregates them jointly, incurring high latency without reusing past computation. The recurrent, state-propagation paradigm (b) propagates the previous state (Ft−1,Lt−1) frame by frame, enabling streaming inference, but offers no mechanism to prevent corrupted observations from contaminating the hidden state under occlusion. Our memory-maintained, contamination-aware paradigm (c) maintains a video memory bank of historical features F , obstacle masks M , and lane masks L . After lane-guided affine alignment, two paths process the aligned data: SSM-Consistent Dual-Gating for occlusion-aware temporal filtering and SSR for lane-guided spatial recovery.
Figure 2: Overview of the SIGMA-Lane Aggregation Module. Video Memory stores historical features Ft−T+1:t , obstacle masks Mt−T+1:t , and the previous lane mask Lt−1 . The Affine Warp module aligns historical data to current-frame coordinates. Warped features then pass through two paths: (1) SSM-Consistent Dual-Gating with a scale-pyramid design for multi-scale temporal propagation and occlusion-aware gating, (2) SSR where current features Ft retrieve spatial context from warped Ft−1 enhanced by Lt−1 embeddings. The outputs are aggregated with current-frame features Ft via Add&Norm and FFN blocks, producing enhanced features F~t .
Figure 3: SSM-Consistent Dual-Gating with a scale-pyramid design. Input features X pass through LN and are split into high- and low-resolution branches. Each branch applies Input Gate, concatenates the Geometry-Aware Start Token s0 , performs Mamba2 [ 4 ] temporal modeling, and applies Output Gate, with both gates modulated by the obstacle mask M . The low-resolution branch is upsampled and merged with the high-resolution branch through Add&Norm to produce X~ . Finally, the current-frame features x~t are extracted from X~ and sent to subsequent modules.
Figure 4: Lane-Guided Affine Warp and Structural Spatial Retrieval. (a) In Lane-Guided Affine Warp, features xt−1 and lane mask ℓt−1 are fed into Lane-Masked GAP to compute lane-region-weighted statistics. Combined with current-frame features xt , an MLP predicts affine parameters p , where tanh constrains the transformation magnitude. The Affine Warp operator synchronously transforms xt−1 , ℓt−1 , and mt−1 using the same grid, yielding aligned outputs x~t−1 , ℓ~t−1 , and m~t−1 . (b) In Structural Spatial Retrieval, current-frame features xt pass through LN to form the query Q . Warped features x~t−1 and lane-mask embedding e(ℓ~t−1) are normalized via LN and summed to form key K and value V . Multi-head cross-attention produces the enhanced feature output.
Figure 5: Visualization of the spatial prior learned by Geometry-Aware State Initialization. Each strip shows the input image, position-prior heatmap, and ground-truth lanes. The learned prior is concentrated around lane-like structures and provides a cold-start bias before reliable temporal history is available.