Practical overview
When it fits
Use it when you need DINOv2 implementation code or pretrained checkpoints oriented toward remote-sensing images, or when its evaluation scripts and reported dataset results are relevant to your experiment.
Who it is for
- Remote-sensing researchers and developers comparing pretrained vision-transformer models on image datasets.
- Practitioners adapting DINOv2 training or evaluation code for aerial and satellite imagery.
Documented uses
- Select a pretrained small, base or large model for an experiment.
- Evaluate a model with the repository’s linear-probing or KNN scripts.
- Inspect a DeepSpeed ZeRO Stage 3 multi-GPU training setup.
Requirements & limits
- Check the checkpoint scale: the README lists ViT-S, ViT-B and ViT-L variants.
- The repository documents PyTorch, Hugging Face Transformers and DeepSpeed; training uses DeepSpeed ZeRO Stage 3.
- The README lists Million-AID and SkyScript as training datasets and nine datasets for evaluation.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Run the documented linear-probing evaluation
Evaluate a model on a dataset using the repository’s linear-probing script.
Before you start
- A model path for the evaluation script.
- A dataset root.
- Training and test text-file lists; the README provides example paths for RESISC.
Steps
Set the model path and dataset root.
Provide the training and test text-file paths.
Run evaluation/linprob.py with the --model-path, --data-root, --train-text and --test-text arguments.
Expected outcome
The documented procedure evaluates the model with linear probing on the selected dataset split.
Guide limitations
- The README gives argument placeholders rather than a complete dataset setup procedure.
- The reported evaluation uses a random 90/10 training-test split.
Set up multi-GPU training with DeepSpeed ZeRO Stage 3
Use the repository’s documented DeepSpeed training invocation with a model path and batch size.
Before you start
- DeepSpeed and the repository’s training code.
- A model path and total batch size.
- The ds_config/ds_z3.json configuration file.
- The README’s example specifies localhost GPU indices 0 through 7.
Steps
Choose the model path and total batch size.
Use the documented DeepSpeed invocation with train.py, the ZeRO Stage 3 configuration, and the --deepspeed flag.
Expected outcome
The documented command specifies multi-GPU training through DeepSpeed ZeRO Stage 3.
Guide limitations
- The README does not provide a full environment installation or dataset-preparation procedure.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- chagmgang/dinov2-remote-sensing · Official homepage title · retained in its source language