Detection, segmentation, depth and recognition research, plus the vision backbones that keep displacing the last generation.
U-Net remains a practical baseline for image segmentation because of its simple encoder-decoder structure and skip connections. However, the bottleneck representation is still dominated by a limited s...
The paper proposes a confidence‑normalized continuous multi‑timescale event representation that encodes temporal information directly into event data using logarithmic B‑spline temporal encoding and a geometry‑aware local confidence mechanism. When applied to a fixed feed‑forward EventCenterNet detector, this representation outperforms the compact CSTR representation on the PEDRo and Gen1 datasets. Additionally, a recursive exponential‑polynomial approximation is introduced to allow efficient event‑by‑event updates while largely preserving detection performance.
The paper introduces SoundMHPE, the first system to estimate multi‑person 3D poses using only acoustic signals. It tackles challenges such as overlapping acoustic signatures and inter‑person reflections by employing an Acoustic Multi‑scale Encoder and a Temporal Pose Decoder with attention. The authors built a 6‑hour Acoustic Multi‑person Pose dataset and show that SoundMHPE outperforms baseline models.
VeriPhy is an auditable physical‑verification system that transforms a text prompt into typed physical obligations and a statically validated execution plan before any video frames are generated. During execution, it gates calls to frozen low‑level experts (segmentation, tracking, counting, depth, OCR, audio‑event detection, etc.) and records provenance‑carrying evidence for each action. The system maps these records to a three‑valued state—supported, contradicted, or unknown—providing traceable verdicts that can be used to refine generation models.
The paper introduces TRIM, a black‑box defense for backdoor attacks in computer vision models. TRIM identifies and removes malicious trigger regions at inference time using region‑based segmentation, adaptive trigger discovery via inpainting and diffusion, and selective purification, without needing model internals, training data, or clean samples. Experiments on various datasets and trigger types show TRIM reduces attack success rates to as low as 1.16% while maintaining high clean accuracy.
Jina-OCR-v1 is an end‑to‑end document parsing model designed for low‑budget GPUs, combining a compressed‑vision encoder with a 3B mixture‑of‑experts decoder that activates about 570 M parameters per token. It uses a FastMTP speculative decoding head that shares a single draft block across three prediction steps, with greedy verification ensuring lossless decoding. Post‑training includes instruction alignment, robustness fine‑tuning on difficult documents, and GRPO with dense verifiable rewards, achieving 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR‑Bench while delivering the highest page throughput at 2.57 pages per second on an NVIDIA L4 GPU.
The paper presents MaxBoxCount, the winning solution to the iWildCam 2021 Challenge, which tackles counting animals in camera‑trap image sequences without using count labels. It combines a robust species classification pipeline with a counting heuristic based on MegaDetector detections to estimate the number of unique individuals across short image bursts. The method addresses challenges posed by temporal discontinuities and the high cost of manual count annotations.
WireSeg-32K is a new synthetic dataset designed for wire instance segmentation, comprising 32,000 RGB images, instance masks, depth maps, and a complementary real-world test set with annotations. The dataset is generated using DeformX, a co-simulation pipeline that couples Cosserat-rod dynamics with photorealistic Isaac Sim rendering to produce physically plausible, contact-consistent wire shapes and diverse scenes. Fine-tuning SAM3 with LoRA on WireSeg-32K improves real-world mAP@75 by 10.2% compared to the off-the-shelf model, demonstrating the transferability of physically grounded synthetic data.
Exemplar is a few‑shot segmentation method that combines a frozen DINOv3 backbone with a fixed bank of classical native‑resolution filter responses in a single lightweight head. Trained only from support masks, it achieves a mean foreground intersection‑over‑union of 0.782 across eleven biomedical imaging datasets, outperforming either component alone and surpassing five other few‑shot methods in 54 of 55 comparisons. With a single annotated mask, Exemplar reaches 0.703, higher than a from‑scratch nnU‑Net trained on the same mask, and while nnU‑Net eventually overtakes it with eight masks, it requires 16–77× longer to fit.
The paper introduces FlexibleFusion, a method for infrared-visible object detection that adapts to both complete and missing-modality scenarios. It employs a Modality-Aware Experts Collaboration mechanism to selectively fuse cross-modal or intra-modal pathways, and a Residual Self-Paced Entropic Optimal Transport module to align heterogeneous feature distributions without heavy optimization. Experiments demonstrate consistent performance across various modality configurations.
The paper explores how synthetic data can be used to train a Thai OCR model without real OCR labels. By systematically varying factors such as typeface diversity, page structure, and handwriting glyphs, the authors identify which aspects of realism most improve transfer to real documents. Using these insights, they adapt a large PaddleOCR model into Wayu‑Paxa‑OCR‑Zero, achieving a median character error rate of 1.24% on printed pages and 20.55% on handwriting, outperforming existing Thai OCR systems.
The paper introduces a unified bundle adjustment framework that jointly optimizes camera parameters, sparse 3D points, and richer geometric features such as lines, coplanarity, and parallelism. It classifies features into scalable ones with direct 2D measurements and higher‑order groups that can be treated as camera‑like entities, allowing group constraints and cross‑feature relations to be expressed via 2D reprojection errors. This approach preserves the sparsity of classical point‑based BA, maintains numerical stability, and achieves runtime comparable to point‑only BA while producing richer 3D structures and improved accuracy.
AnyBox is a zero‑shot framework that estimates the full 9DoF pose (6D pose plus 3D dimensions) of boxes from a single RGB‑D image, leveraging the geometric regularity of boxes. It alternates between pose and scale estimation, using a binary search guided by the discrepancy between a reprojected template and the observed mask, and employs a depth‑consistency filter and an early‑stopping rule to prune implausible hypotheses. On public benchmarks and a warehouse dataset, AnyBox improves detection AP by up to 36 points and boosts robotic box‑shelving success by 28%.
The paper introduces XCal-FL, a federated learning algorithm that dynamically calibrates differential privacy noise using three signals—prediction logit variations, counterfactual margins, and saliency concentration—to improve both predictive accuracy and explanation fidelity. Experiments on medical imaging datasets demonstrate that XCal-FL outperforms static-noise and state‑of‑the‑art adaptive DP methods, achieving over 10% better accuracy and up to fivefold higher explanation fidelity while using privacy budgets more efficiently. The study highlights that explanation fidelity behaves non‑linearly with privacy loss, indicating that explainability is a separate dimension of the privacy trade‑off.
The paper introduces set difference captioning for autonomous driving datasets, aiming to generate natural‑language descriptions of differences between two image subsets. It adapts a two‑stage approach to focus on object‑centric patches, allowing attribution of differences to specific objects or categories. A new benchmark, AD‑Diff Bench, is presented to evaluate these methods, especially for sparse, real‑world differences, with open‑weight models to ensure reproducibility.
The paper presents a method to replace traditional window-based tracking in geostationary atmospheric motion vector (AMV) stereo matching with deep optical flow, enabling efficient and accurate retrieval of dense three‑dimensional wind fields. A stereo teacher model is distilled into a single‑satellite student model that emulates the teacher’s uncertainty estimates, allowing global wind generation from full‑disk GEO imagery. Validation against radiosondes, operational AMVs, ERA5 reanalysis, and EarthCARE cloud profiles shows that the stereo winds outperform operational AMVs in water‑vapor bands while performing slightly worse in the long‑wave infrared band.
ENEAS is a unified, text‑promptable method that simultaneously provides precise instance tracking and high‑quality segmentation, and enables open‑concept discovery of any instance named by a text query. It extends the SeC architecture with a text‑prompting adapter and temporal memory to maintain targets through disappearance and avoid drifting, while a semantic verification layer combines visual embedding matching with conditional VLM refinement to filter ontological errors. Designed for 3D reconstruction, ENEAS delivers robust semantic tracking and segmentation across videos, libraries, and unordered collections, distinguishing true instances from look‑alike doppelgangers.
The study explores how adding anatomical priors and active learning can improve the accuracy of deep learning models for segmenting the Clinical Target Volume (CTV) in gastric cancer radiotherapy. Using 100 retrospective CT scans, an nnU‑Net model trained on 10 expert‑contoured cases was enhanced with voxel‑wise anatomical prior maps and iterative active learning over four rounds. The combined approach raised the mean Dice Similarity Coefficient from 0.84 to 0.87, demonstrating that both techniques individually and together improve segmentation performance and generalizability.
The paper investigates using manufacturer catalogue photography to bootstrap a computer‑vision system for recognizing carbide rotary burrs, a task that traditionally relies on manual quality checks. It shows that while frozen feature extractors struggle to separate key attributes like head shape and tooth profile, metric learning can cluster catalogue images almost perfectly, yet only about half of this performance transfers to real field photographs. The study finds that simple domain‑sensitivity reductions—such as converting images to grayscale and applying Hungarian assignment based on order sheets—yield the largest gains, suggesting catalogue images are a useful cold‑start source rather than a ready‑for‑deployment training set.
ARCOS is a patch-based zero-shot boundary localization framework designed for corneal layer segmentation in optical coherence tomography (OCT) images. It predicts boundary heatmaps from overlapping native-resolution patches, stitches these predictions across full B-scans, and refines them with a self-conditioned module to produce continuous, anatomically ordered layer segmentations. Evaluated on clinical OCT data from multiple devices, ARCOS achieved 95.1% off‑by‑one boundary accuracy on matched-device tests and 84.3% on unseen devices, with low mean absolute boundary errors and accurate thickness estimates.