AI-powered multi-image co-segmentation using DINOv2 and Segment Anything Model (SAM) with Semantic Consensus Prompting.
Linear probing of patch-level representations from ViT-based models (CLIP, DINO, MAE...) on a semantic segmentation dataset.
This is a warehouse for DinoV2-models, based pytorch framework.
Hands-on 10-chapter clinical AI curriculum for radiology fellows with zero ML background. Notebook + glossary + Gradio demo + fine-tuned HF model. Designed to be executed, not read.
Feature Island Induction From a Semantically Rich Ocean. Not the title of a deeply philosophical book about representation.
Team entry, 2025 NSF HDR anomaly detection challenge: gravitational wave, butterfly, sea level
Commercial R&D: Lightweight deepfake detection on edge devices. Comparative benchmark of MobileNetV4 (9.2M) vs DINOv2 (22.1M). Custom dataset AuthorDeepFake-6600 (3,300/3,300). Grad-CAM interpretability. Industrial deployment target.
Code for my Master's Thesis "Deep Neural Encoding Models of the Human Visual Cortex to Predict fMRI Responses to Natural Visual Scenes" and my submission for the "Algonauts Project 2023 Challenge".
A high-performance Visual Product Recognition system for identifying products from images using a coarse-to-fine visual retrieval pipeline
Config-driven fine-tuning and benchmarking of Vision Transformers (ViT, DINOv2, Swin, ConvNeXt) against a ResNet-50 baseline for 40-class scene recognition on Places2, with attention-map explainability.
Estimating latitude and longitude with no GPS, by matching a live camera against onboard satellite imagery on an 8 GB edge GPU. ROS 2 / Isaac ROS / DINOv2 / STHN.
Visual maze navigation system using DINOv2, VLAD, ORB and FAISS — Robot Vision course projects at NYU Tandon
GNSS-denied absolute visual localization for UAVs — fuses visual odometry with satellite-map matching in a particle filter. Evaluated on 9 real survey flights (406-2572 m). Runs in 1.4 GB VRAM.
Camera-based chess board reader — a CV system that watches a live over-the-board game and streams timestamped piece placements with per-square confidences. Positions, not moves.
Parameter-efficient adaptation of a frozen DINOv2 ViT-S/14 for face recognition: 94.25% LFW with 238,080 trainable parameters (40,960 in the adapter), trained on a single 8GB GPU
PyTorch training engine for urban perception modeli 10000 ng. Uses ViT multi-head architecture to learn human preferences from pairwise comparisons, with TrueSkill-regularized ranking loss and Optuna HPO.
[GCPR 2026] Official Implementation for MixMatchDet: Mixing and Matching Tokens for Plain ViT Detectors
The official source code of paper in ECCV 2024. "Aligning Neuronal Coding of Dynamic Visual Scenes with Foundation Vision Models"
A deep learning framework for cross-domain plant species identification using DINOv2 Vision Transformers, LoRA fine-tuning, and Multi-Task Learning.
Add a description, image, and links to the dinov2 topic page so that developers can more easily learn about it.
To associate your repository with the dinov2 topic, visit your repo's landing page and select "manage topics."