Title: LoCAtion: A Long-time Collaborative Attention Framework for High Dynamic Range Video Reconstruction

Abstract:

Current approaches to High Dynamic Range (HDR) video reconstruction are largely confined to a delicate alignment-and-fusion paradigm. Although explicit spatial alignment effectively restores fine details in controlled settings, it proves to be a significant limitation in unstructured, dynamic environments. By imposing rigid alignment on unpredictable movements and fluctuating exposures, these techniques inevitably convert registration inaccuracies into pronounced ghosting artifacts and temporal flickering. This study challenges this traditional assumption. Acknowledging that explicit alignment is inherently susceptible to real-world complexities, we introduce LoCAtion, a Long-time Collaborative Attention framework. This approach redefines HDR video generation, shifting it from a fragile spatial warping operation to a robust, alignment-free collaborative feature routing process.

Guided by this novel formulation, our architecture explicitly separates the highly entangled reconstruction task. Instead of attempting to rigidly warp adjacent frames, the system anchors the scene to a continuous medium-exposure backbone. It then employs collaborative attention to dynamically extract and inject reliable irradiance cues from unaligned exposures. Additionally, we present a learned global sequence solver. By utilizing bidirectional context and long-range temporal modeling, this component propagates corrective signals and structural features throughout the entire sequence, thereby naturally enforcing whole-video coherence and removing jitter. Comprehensive experiments reveal that LoCAtion delivers state-of-the-art visual quality and temporal stability, providing a highly competitive equilibrium between accuracy and computational efficiency.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC