Research map / Video / Video segmentation and tracking

E4: Energy-Efficient DNN Inference for Edge Video Analytics via Early Exiting and DVFS

AAAI 2025 · Video › Video segmentation and tracking · approach: Video restoration with temporal priors · 12 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes E4, an early-exit DNN framework for edge video analytics that uses an attention-based cascade to pick exit points per frame complexity and a coordinate-descent JIT profiler to tune CPU/GPU DVFS frequencies per layer.

Problem

Energy-efficient DNN inference for video analytics on resource-constrained edge devices, where existing DVFS governors and early-exit methods ignore video frame complexity and optimize latency/accuracy rather than energy.

Approach

Combine an attention-based cascade of early-exit gates driven by accumulated (LSTM-pooled) frame features with a coordinate-descent search over per-layer CPU and GPU clock frequencies, so exit points and DVFS settings adapt to each video's complexity.

Key results

• E4 achieves 20%-37% energy saving and 1.3x-2.2x speedup over baselines with EfficientNet-B0 across five Jetson devices.
• On Jetson AGX Orin, E4 reduces energy consumption by up to 46% and latency by up to 73% versus the baseline without DVFS or early exit.
• E4 reduces memory usage by up to 55% relative to the baseline and 23% relative to EENet.

Measured effects reported in the paper

Surprising findings

• E4 yields larger gains on higher-compute edge devices (AGX Orin gains 37% higher on average than Nano), attributed to a larger frequency search space.
• DVFS alone increases latency by about 12% on average while reducing energy by 26%.

Limitations

• Gains on lower-compute devices like Jetson Nano are smaller due to a narrower frequency range.
• Using CNNs limits exploitation of temporal information, which caps accuracy gains from more frames.

Experimental setup

Datasets: ActivityNet-v1.3 Mini-Kinetics

Benchmarks: ActivityNet-v1.3 Mini-Kinetics

Metrics: top-1 accuracy mAP energy consumption (W) inference latency (ms) memory usage

Baselines: EENet zTT Ring-DVFS E4-R (random search variant)

Builds on: EENet zTT BranchNet Remix

Closest papers by meaning

• E2-Train: Training State-of-the-art CNNs with Over 80% Energy Savings (NeurIPS 2019 · similarity 0.85)
• The Hidden Joules: Evaluating the Energy Consumption of Vision Backbones for Progress Towards More Efficient Model Inference (ICML 2025 · similarity 0.84)
• Stepping Forward on the Last Mile (NeurIPS 2024 · similarity 0.82)
• Auto-reconfiguration for Latency Minimization in CPU-based DNN Serving (ICML 2025 · similarity 0.82)
• Multi-Edge Reinforced Collaborative Data Acquisition for Continuous Video Analytics by Prioritizing Quality over Quantity (AAAI 2025 · similarity 0.82)
• MEET: Towards Memory-Efficient Temporal Sparse Deep Neural Networks (CVPR 2025 · similarity 0.82)