Rectifying and Consolidating Latent Diffusion for Remote Sensing Object Detection
Abstract
Remote sensing object detection is fundamentally challenged by scale variations and background clutter. Existing discriminative methods often lack the generative comprehension needed to interpret ambiguous objects from incomplete visual cues. While adapting text-to-image diffusion models offers rich semantic priors to bridge this gap, their application is limited by a dual dissonance: semantic manifold misalignment and generative temporal trajectory underutilization. In this paper, we propose rectified consolidated latent diffusion (RCLD), a novel generative-perception framework. Firstly, we present a semantic manifold rectification to purify the semantic space. It uses a learnable latent probe to dynamically filter out irrelevant text embeddings, thereby restricting multi-modal interactions to the valid scene subspace. Then we propose a temporal trajectory consolidation to integrate the generative history. This module aligns the detailed textures of late diffusion stages with the stable categorical certainty established in early stages. Extensive experiments on four challenging benchmarks demonstrate that RCLD achieves state-of-the-art performance, exhibiting superior stability against complex degradations in aerial imagery.