Demo images

July 11, 2026 · View on GitHub

Place test images here (e.g. .jpg, .jpeg, .png) to run inference with odet image-demo or make demo.

Getting demo images

Sample image: Download at least one example into demo/ (top-level files only are picked up by directory mode):
```
curl -sL -o demo/demo.jpg https://raw.githubusercontent.com/open-mmlab/mmrotate/main/demo/demo.jpg
```
Add more .jpg / .jpeg / .png files in demo/ as needed; make demo processes all of them.
Your own images: Use any aerial/satellite or oriented-object image. DOTA-style tiled crops (e.g. 1024×1024) work well if your config was trained on similar data. Use another folder: make demo DEMO_DIR=/path/to/images.

Usage

From the oriented-det project root:

# Single image
odet image-demo demo/demo.jpg runs/<model>/<timestamp>/config.json runs/.../checkpoints/checkpoint_best.pth --out-file result.jpg

# All top-level images in demo/ (outputs to demo/out by default)
odet image-demo demo configs/.../config.json runs/.../checkpoints/best.pth --out-dir demo/out

Makefile (repo root): uses the latest experiment under runs/ (latest checkpoint under runs/):

make demo              # all *.jpg / *.jpeg / *.png directly under demo/ → demo/out/<stem>_detections.png
# Optional: DEMO_DIR=other/dir IMAGE_DEMO_DEVICE=cpu IMAGE_DEMO_OUT_DIR=/tmp/demo-out

Only images directly inside DEMO_DIR are used (not subfolders such as demo/out/). See tools/README.md for more options (--score-thr, --nms-thr, --classes, --zoom, --device, --overlap-pixels / --overlap-ratio). If the image size matches the model input from your config (e.g. 1024×1024), inference uses a single forward pass; otherwise images are padded or tiled to that size. With --zoom 2 or --zoom 4, inference runs on the zoomed image and the final visualization remains at the original size.