Researchers present SpermYOLO, a coordinated YOLO-based detection model aimed at locating sperm cells in microscopic images for computer-assisted semen analysis. The work targets difficulties such as densely packed cells, visually similar artifacts, and sperm-like impurities that complicate automated detection. The paper is an arXiv preprint and reports on detection accuracy and efficiency.
Researchers present S3-Tracker, a self-supervised method for tracking points in endoscopic surgical video, a task needed for aligning live footage with preoperative images during robot-assisted procedures. The approach uses contrastive random walks to learn tracking without manual annotations, aiming to stay reliable under soft-tissue deformation. The work targets computer-assisted intervention and autonomous robotic surgery.
A new arXiv preprint introduces a method for semantic image communication that uses channel-adaptive region adjacency graphs as carriers instead of dense latent tensors or grid-aligned semantic layouts. The approach aims to explicitly encode relationships between image regions so that task-relevant scene structure survives tight channel budgets. The abstract frames the work as addressing a gap in how region-level relations are represented under limited bandwidth.
A new arXiv preprint introduces an instance-aware graph modeling approach for detecting and classifying cells in pathology images. The method aims to capture complex cellular interactions within the tumor microenvironment rather than relying only on visual appearance. Accurate cell detection matters for diagnostic accuracy and treatment planning.
A new arXiv paper proposes an unsupervised approach to learning body keypoints for real-time fall detection, aimed at monitoring older adults in home and clinical settings. The authors compare their method against alternatives under realistic conditions and add predictive bandwidth reduction so that continuous video monitoring uses less data. The work targets a known gap: sustained in-person supervision is hard to maintain, while video streams must be practical to transmit.
A new arXiv preprint proposes an explainable screening approach for Parkinson's disease based on hand-drawn spirals and meanders. The work argues that tremor-driven oscillations, irregular strokes, and unstable curvature in these drawings reveal early neuromotor impairment. It then connects visual attribution methods to clinical reasoning so the model's outputs can be interpreted by clinicians.
A new arXiv paper introduces WAVIE, a lightweight approach to detecting deepfake faces that combines wavelet-based augmentation with intermediate vision embeddings. The authors target a common weakness in existing detectors, which often lose accuracy when they encounter manipulation techniques not seen during training. The work is positioned as a step toward more reliable face forgery detection in real-world settings.
A new arXiv paper introduces FICAug, a method that combines feature-informed clustering with data augmentation to improve facial-expression-based screening for Parkinson's disease. The approach targets the problem of small clinical datasets, which limits how well such screening models generalize. It is presented as an updated preprint on arXiv (2409.17685v3) in the cs.AI and cs.LG categories.
A new arXiv preprint proposes combining classical template matching with variational techniques and statistical fusion to detect anomalies in patterned images. The authors argue that deep learning is often too costly or impractical when training data is scarce, while traditional template matching is interpretable but brittle to changes in scale and geometry. The method aims to keep the simplicity of template-based approaches while improving robustness to such variations.
A new arXiv paper compares near-infrared and standard color imaging for daytime farm-machinery traversability using a site-disjoint evaluation that removes spatial data leakage. The authors find NIR does not consistently beat color once frames from the same locations are kept out of both training and test splits, and note that earlier benchmarks reporting an NIR edge used sequence-level splits. The work suggests sensor choices for agricultural perception should be revisited on properly separated data.
Researchers propose BGM2Pose, a method that estimates a person's 3D pose without physical contact by using ordinary music playing nearby as an active sensing signal. The approach aims to avoid the intrusive chirp signals and impractical setups used by earlier acoustic pose-estimation systems. A revised version of the paper is available on arXiv.
A new arXiv preprint argues that treating video gesture recognition as per-frame classification works for control tasks but is not precise enough for synchronization, such as musical or timing-critical interaction. The author proposes a method based on projected winding order that determines gesture state in step with the video frames rather than after a labeling delay. The work appears in the cs.LG cross-list.
A new arXiv paper treats input resolution as a tunable design choice rather than a fixed setting when identifying distant birds, which appears only tens of pixels wide, to reduce bird strikes at wind farms. The authors run a factorial experiment across 14 image side lengths and six neural network architectures, recording latency directly on an edge device. The work reports how accuracy and on-device inference cost trade off across those configurations.
A new arXiv paper addresses computational cost and multi-scale feature extraction limits in bird's-eye-view (BEV) 3D object detection for autonomous driving. The authors combine bi-level routing with sparse spatial attention to improve the efficiency of dense 2D-to-BEV view transformation. The work targets multi-view camera setups used in self-driving perception.
A project showcased on Hacker News explores garments printed with patterns designed to confuse automated face detection systems. The approach builds on research into adversarial examples, transferring techniques that fool image classifiers onto clothing and accessories. It raises questions about how effectively such countermeasures hold up as surveillance models are retrained and improved.
A new arXiv preprint examines how well synthetic images help binary classification tasks where positive examples are scarce, as in medical imaging and industrial inspection. The authors propose measuring a "discriminative span" and reconstructing a classifier to predict how useful generated samples will be. The work aims to guide synthetic data selection in severely imbalanced settings.
A new arXiv preprint introduces a hybrid ensemble selection framework for multiclass cervical cytology image classification that weighs discriminative performance alongside calibration and selective prediction. The authors argue that raw accuracy is not enough for clinical image analysis, and that the method is meant to deliver trustworthy confidence scores and uncertainty flags so users know when to distrust a prediction. The work appears in the cs.AI category as an early-stage research contribution.
A new arXiv paper applies hyperbolic geometry to open-world object detection in satellite and aerial imagery. The work targets the fact that remote-sensing object categories carry hidden hierarchical structure, which standard Euclidean embedding spaces struggle to represent. The method is designed to flag unknown objects and incrementally absorb them into the model once annotations become available.
A new arXiv paper proposes statistical color matching as a simple, low-risk technique for keeping medical image classifiers accurate when deployment conditions diverge from training, such as hardware differences, color variation, and changing patient demographics. The authors argue that common fixes like color jittering do not provide enough diversity, and reposition this overlooked method as a more sustainable path to domain generalization.
A newly updated arXiv paper introduces a classification approach that combines spatial and frequency-domain representations, with particular emphasis on phase information, to better capture structural relationships between visually similar images. The method targets few-shot fine-grained classification, where a model must distinguish closely related categories using only a small number of labeled examples.
A newly posted arXiv paper, cross-listed in computational linguistics and machine learning, presents Vague2Detect, a detection approach designed to work with vague or functional language prompts. The authors note that fixed-class detectors like YOLO and even open-vocabulary systems such as YOLO-World often misinterpret ambiguous wording and fail to match it to the intended objects. The proposed knowledge-based method aims to close this gap for real-world detection scenarios.
SloMoDeblur is a large-scale dataset designed to advance motion blur removal in smartphone photography. Its creators argue that current deblurring benchmarks fall short in size, resolution, and relevance to real-world mobile imaging conditions. The paper has been posted on arXiv and cross-listed under the AI and machine learning categories.
A new arXiv preprint presents a latent bridge matching technique for estimating albedo, the reflectance component separated from lighting in intrinsic image decomposition. The authors note that generative approaches to this task have struggled with weak physical plausibility and computationally heavy inference, and their method is designed to tackle both shortcomings. The work is cross-listed under arXiv's artificial intelligence category.
An arXiv paper (cross-listed to cs.LG) studies long-horizon UAV control driven by camera input, addressing problems like unreliable policy training, overly risky exploration behavior, and the heavy compute cost of processing raw imagery. The approach relies on compact visual cues centered on the target to make learning more stable and practical.
Researchers describe a deep learning method for remote photoplethysmography (rPPG), which estimates heart rate from ordinary video without any physical contact, a capability aimed at telemedicine use. The approach relies on pseudo-labeling to lessen the need for manually annotated training data, and the authors report state-of-the-art results for contactless heart-rate estimation.
A new arXiv paper frames novel-view synthesis with 3D Gaussian Splatting as a structured regression problem and applies conformal prediction to it. The goal is to give rendered views formal statistical coverage guarantees, going beyond uncertainty heatmaps that offer no such certification. The approach is aimed at settings where rendered outputs must meet verifiable reliability standards.
A new arXiv paper introduces Elastoformer, a method that reshapes transformer models at runtime using elastic transformations so they can adjust to shifting operating conditions. The work targets computer vision workloads on edge devices, where latency, power, and compute budgets fluctuate. By adapting model capacity on the fly, the approach aims to let on-device systems balance accuracy against resource constraints in real time.
A new arXiv preprint benchmarks several pre-trained convolutional neural networks on the task of separating melanoma from other skin lesions. The authors frame the problem around the difficulty of visual similarity between lesion types and variability in imaging, which complicates early diagnosis. The comparison evaluates how well these existing image models transfer to this clinical classification setting.
A new arXiv preprint describes a deep-learning pipeline that boosts detection of dim moving objects in optical space situational awareness imagery. The method combines automated star removal with background reconstruction to help recover objects at low signal-to-noise ratios. The authors frame the work as a step toward more reliable tracking of faint orbital targets.
An arXiv paper describes a submission to the EgoProactive track of the ECCV 2026 Wearable AI Challenge, which placed first in the large-model division and second in one other division. The work focuses on proactive egocentric assistance, meaning systems that anticipate a wearer's needs from first-person visual input, and uses visually grounded supervision to train the model. The decision is framed as a yes/no prediction derived from renormalized scores.
A paper summarizes the FedSurg EndoVis 2024 Challenge, which tested federated learning methods on surgical video for appendicitis classification without centralizing patient data. The work addresses the difficulty of building generalizable surgical AI when hospitals cannot share video directly, and reports benchmark outcomes from participating teams. It positions federated training as a viable approach for privacy-sensitive, spatiotemporal surgical tasks.
A revised arXiv preprint describes a method that combines deep feature extraction techniques to identify surface defects on solar panels. The authors motivate the work by noting that manual inspection of large-scale solar plants is labor-intensive, slow, and expensive. The approach aims to support automated monitoring for plant efficiency and reliability.
A new arXiv paper presents an approach to automated wound care that combines dedicated instance segmentation models for detecting wound boundaries with multi-class classification. The authors argue that existing AI systems for wound analysis tend to be narrow in scope, and propose handling boundary detection and wound typing as separate, specialized tasks. The method targets clinical decision support in both chronic and acute wound management.