[2510.13714] DeDelayed: Deleting Remote Inference Delay via On-Device Correction

Electrical Engineering and Systems Science > Image and Video Processing arXiv:2510.13714 (eess) [Submitted on 15 Oct 2025 (v1), last revised 2 Apr 2026 (this version, v3)] Title:DeDelayed: Deleting Remote Inference Delay via On-Device Correction Authors:Dan Jacobellis, Mateen Ulhaq, Fabien Racapé, Hyomin Choi, Neeraja J. Yadwadkar View a PDF of the paper titled DeDelayed: Deleting Remote Inference Delay via On-Device Correction, by Dan Jacobellis and 4 other authors View PDF HTML (experimental) Abstract:Video comprises the vast majority of bits that are generated daily, and is the primary signal driving current innovations in robotics, remote sensing, and wearable technology. Yet, the most powerful video understanding models are too expensive for the resource-constrained platforms used in these applications. One approach is to offload inference to the cloud; this gives access to GPUs capable of processing high-resolution videos in real time. But even with reliable, high-bandwidth communication channels, the combined latency of video encoding, model inference, and round-trip communication prohibits use for certain real-time applications. The alternative is to use fully local inference; but this places extreme constraints on computational and power costs, requiring smaller models and lower resolution, leading to degraded accuracy. To address these challenges, we propose Dedelayed, a real-time inference system that divides computation between a remote model operating on delay...