alchemistUserhf/itapon-yolo

iTapon YOLO-World (v0, zero-shot)

The waste detector behind iTapon, a mobile-first web app that tells users which bin an item goes in (Biodegradable / Recyclable / Residual / Hazardous, following the Philippine RA 9003 categories). It's a CS 406 Machine Learning project at the Technological Institute of the Philippines, Quezon City.

v0 is zero-shot, not trained on our own data. It is Ultralytics yolov8s-worldv2 with the 23 iTapon classes baked in as 46 text prompts (custom vocabulary), exported to ONNX for in-browser inference. A model trained on a local dataset is planned and will replace it under a new tag.

Files

File	Size	Use
itapon-world-v0-480-int8.onnx	13.2 MB	Default for the web app (dynamic INT8 quantization, 480 px input)
itapon-world-v0-640-int8.onnx	13.3 MB	Option for strong devices / kiosks (640 px input: slower, higher scores)
itapon-world-v0.labels.json	5 KB	Output class index → class_key, default_category
SHA256SUMS	—	Checksums

Input / output

• Input images: float32 [1, 3, 480, 480] (or 640 for the 640 file), RGB scaled 0–1, letterboxed (aspect ratio kept, gray 114 padding, centered).
• Output: float32 [1, 10, 6], one row per detection: x1, y1, x2, y2 (input pixels), score, class_index. Class-agnostic NMS is built in (IoU 0.5, score ≥ 0.10). Rows with lower scores are padding.
• Map class_index through labels.json. Several prompts map to the same class_key.

Classes (23)

Recyclable: plastic_bottle, plastic_cup, aluminum_can, glass_bottle, paper_bag, cardboard, paper_sheet · Biodegradable: banana_peel, food_scraps, leaves, used_tissue · Residual: snack_wrapper, styrofoam, plastic_utensil, straw, tetra_pack, soiled_food_container, sanitary_napkin, face_mask · Hazardous: battery, light_bulb, expired_medicine, spray_can

The bin shown to users comes from the app's database, not from this model, so local rules can differ without retraining.

Usage (browser)

import * as ort from "onnxruntime-web/wasm";
const session = await ort.InferenceSession.create(
  "https://huggingface.co/alchemistUserhf/itapon-yolo/resolve/v0.1/itapon-world-v0-480-int8.onnx",
);
const out = await session.run({ images: new ort.Tensor("float32", chw, [1, 3, 480, 480]) });

See apps/web/src/lib/detector/ in the repo for the full pre- and post-processing.

Evaluation

There's no accuracy evaluation yet. v0 has only been smoke-tested on a few photos (paper bag and plastic cup detected correctly). Desktop Chrome (WASM, 4 threads): ~420 ms per frame at p50.

Limitations

• Zero-shot accuracy is uneven. Visually similar classes (wrappers, Tetra Pak, clean vs. greasy paper) are weak.
• The model can't judge contamination (greasy vs. clean). The app shows a caveat instead.
• Scores are not calibrated probabilities. The app uses display tiers (≥ 0.35 confident, ≥ 0.15 "possibly").

License

AGPL-3.0, matching Ultralytics YOLO. Source: https://github.com/alchemistUser/itapon