ATHENA Christian Doppler (CD) Laboratory

EVLM: Intent-Driven Edge Vision Language Model for UAV-Based Power Line Inspection

2026 IEEE International Conference on Edge Computing and Communications (IEEE EDGE 2026)

Reza Farahani (TU Wien, Austria), Zoha Azimi (AAU, Austria), Ilir Murturi (University of Prishtina, Kosovo), Arda Goknil (SINTEF, Norway), Sagar Sen (SINTEF, Norway), Christian Timmerer (AAU, Austria), Schahram Dustdar (TU Wien, Austria)

Abstract: Inspection of critical infrastructure, such as power lines, is increasingly conducted using unmanned aerial vehicles (UAVs) that capture aerial video for subsequent human review. Although recent edge-based approaches deploy onboard object detectors to identify predefined defect classes, these pipelines remain closed-set, task-specific, and largely decoupled from operator intent and edge resource constraints. This paper introduces EVLM, an intent-driven vision-language framework for onboard UAV-based power line inspection. Given a high-level operator intent, EVLM (i) leverages lightweight histogram-based frame filtering to extract salient key frames under bounded compute budgets, (ii) executes a domain-adapted vision language model (VLM) directly on the UAV for intent-conditioned multimodal reasoning, and (iii) synthesizes structured inspection reports together with a minimal set of evidence frames, replacing continuous raw video transmission with compact semantic outputs. To align the VLM with infrastructure inspection semantics while preserving edge efficiency, we perform parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA), enabling domain specialization without updating the full model parameters. We implement and fully deploy EVLM on an NVIDIA Jetson device representative of UAV-class onboard hardware and evaluate it using 20 publicly released power line inspection video sequences spanning 8 heterogeneous environments and 5 operational intent categories. Experimental results show a data reduction of 94.8 %, with transmitted data decreasing from 485 kB to 25 kB per 4 s segment, corresponding to 72.75 MB versus 3.75 MB over a 10 min inspection mission. EVLM operates feasibly on embedded hardware, maintaining moderate CPU/GPU utilization and bounded power consumption (5.6 W), while producing interpretable, intent-aligned inspection outputs with richer semantic insights than detection-centric baselines.

Assistant Prof. Dr. Hadi Amirpour has been elevated to IEEE Senior Member in recognition of his contributions to multimedia streaming systems.

IEEE Senior Member is the highest professional grade for which an IEEE member can apply. This distinction requires extensive professional experience and demonstrated accomplishments that reflect technical expertise, leadership, and professional maturity. Fewer than 10% of IEEE’s nearly half a million members worldwide have achieved this honor.

Energy and Compression Efficiency in Large-Scale Video Streaming

IEEE International Conference on Image Processing (ICIP 2026)

[PDF]

Mohammad Ghasempour (AAU, Austria), Hadi Amirpour (AAU, Austria),  and Christian Timmerer (AAU, Austria)

Abstract: The rise in large-scale video streaming has led to increased energy demands across the encoding, transmission, and decoding pipeline. While energy consumption in video streaming has been widely studied, encoding decisions are typically made without explicitly accounting for expected content demand. As a result, the impact of view count on energy consumption and compression efficiency remains largely unexplored. This limits the ability to make informed and efficient encoding decisions in real-world streaming scenarios. In this paper, we propose EcoEncode, an analytical framework to evaluate the impact of view count on codec-level encoding decisions and the resulting trade-offs between energy consumption and compression efficiency. We further show that these decisions depend on video content characteristics and encoding configurations. Based on our findings, we provide practical insights to guide the selection of codecs and presets. Experimental results show that view count is a key factor in codec-level decisions. For low-popularity videos, EcoEncode achieves up to 99% energy savings with only 1-4 VMAF points of quality loss. Across all scenarios, the selected configurations lie on or near the Pareto frontier, and EcoEncode improves quality by up to 14 VMAF points over the least energy-consuming configuration.

As part of the “FÄKT” initiative by the Austrian Academy of Sciences, which produces science videos for audiences aged 10 to 14, a recent episode highlights insights into the world of video streaming: Christian Timmerer, head of the Christian Doppler Laboratory for Adaptive Streaming over HTTP and Emerging Network-based Multimedia Services at the University of Klagenfurt and a two-time Technology & Engineering Emmy® Award recipient, explains how video streaming works on a global scale and how it is continuously optimized.

The short video provides an accessible overview of the technologies behind modern streaming services and demonstrates how research contributes to improving the quality and efficiency of video delivery. Please note that the video is currently available in German only.

Links

• Video
• FÄKT-Website

EPS: Efficient Patch Sampling for Video Overfitting
in Deep Super-Resolution Model Training

IEEE Transactions on Circuits and Systems for Video Technology (TCSVT)

[PDF]

Yiying Wei (AAU, Austria), Hadi Amirpour (AAU, Austria),  Jong Hwan Ko (SKKU, South Korea), and Christian Timmerer (AAU, Austria)

Abstract: Leveraging the overfitting property of deep neural networks (DNNs) is trending in video delivery systems to enhance video quality within bandwidth limits. Existing approaches transmit overfitted super-resolution (SR) model streams for low-resolution (LR) bitstreams, which are used to reconstruct high-resolution (HR) videos at the decoder. Although these approaches show promising results, the huge computational costs of training a large number of video frames limit their practical applications. To overcome this challenge, we propose an efficient patch sampling method named EPS for video SR network overfitting, which identifies the most valuable training patches from video frames.
To this end, we first present two low-complexity Discrete Cosine Transform (DCT)-based spatial-temporal features to measure the complexity score of each patch directly. By analyzing the histogram distribution of these features, we then categorize all possible patches into different clusters and select training patches from the cluster with the highest spatial-temporal information. The number of sampled patches is adaptive based on the video content, addressing the trade-off between training complexity and efficiency.
Our method reduces the number of training patches by 75.00% to 91.69%, depending on the resolution and number of clusters, while preserving high video quality and greatly improving training efficiency. Our method speeds up patch sampling by up to 82.1× compared to the state-of-the-art patch sampling technique (EMT).