Adaptive keyframe selection and information-guided sampling for efficient RGB-D neural SLAM

Accurate 3D reconstruction under strict computational constraints is a central challenge for deploying simultaneous localization and mapping (SLAM) in real-time autonomous systems. Recent neural SLAM methods achieve high reconstruction fidelity through implicit scene representations and hybrid encodings, yet they universally rely on fixed-interval keyframe insertion and uniform random pixel sampling to update the map. This combination leaves a fundamental efficiency gap: the system cannot distinguish informative frames from redundant ones, nor allocate computation to the image regions that would benefit most from further optimization. To close this gap, we propose an adaptive data selection framework that jointly adapts which frames enter the optimization and which pixels within those frames receive computational effort, two dimensions that all prior neural SLAM methods address at most independently. Specifically, the temporal component uses photometric error-based adaptive thresholding with exponential moving average and momentum correction to insert keyframes only when a frame carries novel information relative to the current scene model. The spatial component replaces uniform random sampling with an importance map that fuses gradient magnitude and photometric residuals, directing optimization toward geometrically complex and poorly reconstructed regions. Experiments on four standard RGB-D benchmarks (Replica, ScanNet, TUM RGB-D, and Synthetic RGBD) demonstrate that the proposed framework processes only 20 to 30% of frames and 60% of pixels per iteration while achieving an average tracking accuracy of 8.28 cm ATE on ScanNet (compared with 8.75 cm for the baseline) and 1.25 cm on TUM RGB-D (compared with 2.17 cm), operating at up to 28.5 FPS on a desktop equipped with a 3.60 GHz Intel Core i7-12700K CPU and an NVIDIA GeForce RTX 3090 Ti GPU. These results confirm that principled joint data selection yields up to 2× speedup with a 40% reduction in sampled pixels while maintaining or improving reconstruction quality, making the framework suitable for real-time deployment on resource-constrained hardware.