Multimedia AI now characterizes multiple concurrent streams at high resolutions, spanning video, audio, language, and multimodal fusion. Energy, bandwidth, and even water for cooling have become first-order constraints for real deployments. Training a single large NLP model has been estimated to emit up to 284 tons of CO2, while infrastructure and model choices can shift emissions by orders of magnitude; at the same time, data centers consume huge volumes of water for cooling (hundreds of thousands). In this position paper, we argue that we cannot continue to choose models based on accuracy and throughput alone. We propose a Sustainability Card for Multimedia AI-Correct Outputs per kWh, Joules per Sample, Bytes per Output, and optional Water per 1,000 Outputs-and outline a simple measurement protocol.
We present "Dialogue-Pseudo", a dialogue-aware pseudonymization framework that (i) maps each real speaker to a distinct pseudo-speaker via a dialogue-aware PartitionAssignment-Selection procedure to preserve inter-speaker separability, (ii) replaces personally identifiable information with syllable- and accent-matched surrogates to maintain phonorhythmic timing, and (iii) applies bounded prosody control that normalizes global pitch/energy while preserving relative contours and turn-taking cues. Experiments on the RWCP conversational corpus with a pseudo-speaker pool derived from Common Voice show improved privacy under an ECAPA-TDNN speaker verification model: Equal Error Rate (EER) increases from 37.38% (original) to 42.22% (pseudonymized). Prosody-oriented statistics (F0 mean/variance, RMS energy, pause ratio, turn length) remain close across conditions, and UMAP visualizations indicate that pseudo-speakers are well separated from both the sources and from one another. These results support Dialogue-Pseudo as a practical approach to privacy-preserving dialogue analytics that extends pseudonymization beyond timbre conversion to dialogue-relevant linguistic and prosodic cues, thereby balancing unlinkability with the retention of task-critical information.
Large language models (LLMs) can be used to solve various tasks based on text inputs, e.g., video quality estimation. We explore the usage of LLMs for video quality prediction based on metadata (video codec, bitrate, resolution), which has not been addressed before. For the evaluation we use test #1 from our AVT-VQDB-UHD-1 dataset. We generated text prompts based on the metadata and collected answers from 17 different LLMs. The evaluation indicates that especially larger LLMs could be used to simulate human raters. However, a pure model prediction with one model has lower performance than SoA video quality models. Thus, we further investigated combinations of LLMs, which resulted in comparable performance to state-of-the-art models. Our work is a proof-of-concept, considering that LLMs are slower for the prediction than the traditional metadata-based models.
This study focuses on improving the safety of Vulnerable Road Users (VRUs) in traffic by leveraging multispectral imaging and deep learning. We introduce ScaleFuse, a novel RGB-thermal fusion architecture specifically designed to address the limitations of existing multispectral detection methods. Unlike conventional approaches, ScaleFuse performs multiscale feature fusion at intermediate layers of the backbone, adaptively learning spatial and channel-wise importance from both modalities. To ensure robust fusion, we employ the SuperGlue network for precise image alignment, mitigating the common issue of misregistration between RGB and thermal inputs. ScaleFuse is implemented within the YOLOv10 framework, enabling efficient and accurate detection in challenging conditions such as low-light environments, adverse weather, and occlusions. Experimental evaluations on LLVIP, FLIR, and our newly introduced VeNIT-MSD dataset demonstrate that ScaleFuse consistently outperforms single-modality baselines and previous fusion strategies in terms of accuracy, precision, and recall. The proposed system achieves real-time inference while remaining fully compatible with deployment frameworks such as TensorRT, making it suitable for intelligent transportation applications.
Malaria remains a significant public health challenge, particularly in developing countries. Traditional machine learning and deep learning classification approaches typically rely on large, labeled datasets that include both normal (malariafree) and abnormal (infected) samples. However, in the medical domain, the acquisition of annotated infected samples is difficult due to privacy restrictions and limited availability, which restrict data access and sharing. Anomaly detection techniques offer a compelling alternative by reducing the dependency on labeled abnormal data. In this paper, we propose a one-class autoencoder model trained solely on microscopic images of healthy (normal) blood cells. The model learns the underlying distribution and structural patterns of normal cells, enabling it to detect deviations indicative of malaria infection. Structural similarity (SSIM) is employed to measure reconstruction error, which is then used to classify cells as normal or infected. Experimental results demonstrate that the proposed method achieves high accuracy in detecting malaria-infected cells in blood smear images, highlighting its potential as a scalable and effective solution for automated malaria diagnosis.
