Feature-Level Reliability of Directional-Kernel Richardson–Lucy Deblurring Under Kernel-Length and Direction Controls

Image restoration lies between camera acquisition and geometric estimation, but pixel improvements may not transfer to motion estimates. We evaluated directional-kernel Richardson–Lucy (RL) deblurring under kernel-length and direction controls. The restoration analysis covered 3071 paired GoPro, RealBlur-J, and RealBlur-R images. An exploratory feature analysis used a fixed 155-image subset with Oriented FAST and Rotated BRIEF (ORB), scale-invariant feature transform (SIFT), two geometry models, ten random directions, NAFNet, and Restormer. A separate task analysis used ten red–green–blue plus depth (RGB-D) sequences from the Technical University of Munich (TUM) benchmark, synthetic 20 ms exposures, and fixed RGB-D perspective-n-point odometry. Estimated directions contained information relative to random angles, yet the tested global RL branches remained below Blur Input on average. Changes in sequence-mean absolute trajectory error (ATE) RMSE ranged from +0.004 to +0.051 m for ORB and from −0.008 to +0.036 m for SIFT. Seeds were averaged within each sequence before inference. No tested branch achieved a robust ATE improvement across both detectors. Pixel, raw-feature, normalized-feature, geometry-state, and trajectory endpoints produced different method rankings. These findings motivate endpoint-specific evaluation. The task experiment does not validate naturally blurred long-exposure video, a deployed simultaneous localization and mapping system, or sensor hardware.