Edge-TSR expone una sobrevaloración de referencia del 20-30% para la inferencia del borde

20-30% relative degradation hits every model when you move from static-image evaluation to real-world streaming deployment on an edge device. That is the central finding from the Edge-TSR paper out of a team working on continuous roadside perception.

Benchmark Numbers Lie About Real-World Edge Performance

Conventional benchmark evaluations ignore temporal instability in streaming video, thermal throttling under sustained load, and workload-dependent performance variability. Edge-TSR tested three state-of-the-art baselines and found consistent 20-30% drops in reality versus the lab. I have seen this pattern before: a model looks great on ImageNet but chokes in production. Here the gap is quantified with real hardware.

Edge-TSR Fixes Temporal Inconsistency Without Cloud Offload

Edge-TSR runs on a single NVIDIA Jetson Orin Nano, integrating detection, tracking, fine-grained classification, and a lightweight track-aware temporal stabilization mechanism. The stabilization recovers up to 10.16% classification accuracy over per-frame inference baselines. Overhead is negligible. No cloud offload required.

Sustained Real-Time Inference Under Thermal Constraints

A 55-minute vehicular deployment over a 26 km route demonstrates sustained operation at 16.18 FPS within safe thermal limits. That is continuous, real-time roadside perception on one embedded device. The system jointly characterizes inference quality, latency, throughput, and thermal behavior during long-duration operation - exactly the metrics that matter when your edge sensor runs for hours, not seconds.

Benchmark-centric evaluation creates a false sense of capability for edge AI. Edge-TSR shows that deployment-aware evaluation and temporal inference stabilization are not optional; they are necessary for any continuously operating edge perception system.

Source: Beyond Benchmarks: Continuous Edge Inference for Fine-Grained Roadside Perception
Domain: arxiv.org