README.md

metadata

title: Object Counter
emoji: 🔎
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 5.9.1
app_file: app.py
pinned: false
license: mit
short_description: Count any object in an image by typing what to look for

Object Counter

Count any object in an image — people, cars, sticks, steel rods, or anything else you can name — using open-vocabulary object detection. Type what you want counted at runtime; no fixed class list, no training required.

How it works

This project uses YOLO-World, an open-vocabulary detector. Instead of being limited to a fixed set of categories (like standard YOLO's 80 COCO classes), you give it a comma-separated list of object names at runtime, and it detects and counts each one in your image:

car, person, steel rod, stick

The model draws a bounding box around every instance it finds, and the app tallies the boxes per class.

Repo structure

File	What it is
app.py	Standalone Python script — Gradio app with a text box for object classes. Run locally.
requirements.txt	Dependencies for app.py.
object_counter_gradio_app.ipynb	Same Gradio app, packaged as a Colab notebook (free GPU, no local setup). Recommended starting point.
object_counter_yolo_world.ipynb	Earlier, simpler version with a hardcoded class list — kept as a minimal reference example.

Getting started

Option A — Colab (no local setup, free GPU)

1. Open object_counter_gradio_app.ipynb in Google Colab
2. Enable a GPU runtime: Runtime → Change runtime type → T4 GPU
3. Run all cells (Runtime → Run all)
4. A Gradio link/interface appears at the bottom — upload an image, type the object(s) you want counted (comma-separated for multiple), and hit Submit

Option B — Run locally

pip install -r requirements.txt
python app.py

Open the local URL Gradio prints (usually http://127.0.0.1:7860). Works without a GPU, just slower per image.

Usage

• Type a single object (car) or several (car, person, bicycle, dog) — no fixed list.
• Be as descriptive as helps: "steel rod" detects better than "metal"; "delivery van" is more specific than "vehicle".
• Adjust the confidence slider if results are off — lower catches more objects (with more false positives), higher is stricter.

Tuning results

• Missing objects? Lower the confidence threshold (try 0.05–0.1), or use more descriptive class names.
• Too many false positives? Raise the threshold, or make class names more specific.
• Tightly packed or overlapping objects undercounted (e.g. a bundle of steel rods)? This is a known limitation of box-based detectors — separating touching/overlapping instances into individual boxes is hard. A density-map counting approach (e.g. CountGD), built specifically for packed/repetitive objects, is a planned follow-on for this case.

Known issue: CUDA/CPU device mismatch

On GPU runtimes, model.set_classes() can throw:

RuntimeError: Expected all tensors to be on the same device, but got index is on cpu, different from other tensors on cuda:0

This is a device-handling bug in how YOLO-World's CLIP text encoder gets loaded, not a config error on your end. Both app.py and the Gradio notebook already include the fix: the model is moved to CPU before set_classes() and back to the target device (GPU or CPU) afterward. If you still hit this after pulling the latest version here, restart the runtime/kernel and re-run from the model-loading cell.

Roadmap

• Static image counting (YOLO-World)
• User-defined classes at runtime (no hardcoded list)
• Gradio interface for interactive use
• Standalone Python script version
• Live webcam / video counting
• Density-map counting mode for tightly packed objects

Requirements

Installed automatically via requirements.txt / the notebook's install cell:

• ultralytics
• supervision
• gradio
• opencv-python-headless
• pillow

License

MIT — set in the Spaces config block above. Add a LICENSE file with the full MIT text if you also want it to show up as the repo's license on GitHub. Change the license: field in that block if you'd prefer something else.