Modelpublic

Adit1122/conveyor-belt-recycling-detector

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes9downloads

Model Card

SAM3 + Qwen3-VL Conveyor Belt Recycling Detector

64-class recyclable material detection & counting on conveyor belts, deployed on Jetson AGX Orin 64GB

Architecture

Two-stage pipeline:

1.SAM3 (facebook/sam3) — segments all objects via text-prompted concept segmentation
2.Qwen3-VL (Qwen/Qwen3-VL-2B-Instruct) — classifies each crop into 64 recycling categories

64 Material Classes

Range	Category
0-14	Paper & Cardboard (newspaper, magazines, books, kraft, paperboard, cups, cartons)
15-25	PET #1 & PVC #3 (bottles, thermoformed, clamshells, blister packs)
26-36	HDPE #2 & LDPE #4 (bottles, tubs, pails, film)
37-51	PP #5, PS #6, Other Plastics (bottles, containers, foam, films)
52-60	Metals (aluminum cans/foil/aerosol, steel containers/aerosol)
61-63	Glass & Ceramic (bottles, jars, deposits)

Deployment on Jetson AGX Orin 64GB

Quick Start

bash

git clone https://huggingface.co/Adit1122/conveyor-belt-recycling-detector
cd conveyor-belt-recycling-detector
chmod +x setup_jetson.sh
sudo ./setup_jetson.sh

Performance

Metric	Value
GPU Memory Used	~16 GB / 64 GB
SAM3 Segmentation	200-400ms/frame
VLM Classification	50-100ms/crop
Effective Analysis FPS	1-3 FPS

REST API (port 8080)

bash

curl http://localhost:8080/counts    # Get item counts
curl http://localhost:8080/latest    # Latest detections
curl -X POST http://localhost:8080/detect --data-binary @frame.jpg  # Analyze image

Files

—conveyor_belt_detector.py — Core detection pipeline (works on any GPU)
—jetson_inference_server.py — Jetson-optimized server with camera + REST API
—setup_jetson.sh — Automated setup for JetPack 6.2+
—Dockerfile.jetson — Container deployment option
—config.json — Configuration file
—conveyor_belt_detector.service — systemd service unit

Requirements

—Jetson AGX Orin 64GB with JetPack 6.2+
—USB/CSI/IP camera (1080p overhead view)
—HuggingFace account with SAM3 access (request here)

Models Used

—facebook/sam3 — 860M params, gated
—Qwen/Qwen3-VL-2B-Instruct — open, Apache 2.0

library_name: transformers tags:

—object-detection
—recycling
—conveyor-belt
—sam3
—qwen3-vl
—jetson
—edge-ai

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

—Try ML Intern: https://smolagents-ml-intern.hf.space
—Source code: https://github.com/huggingface/ml-intern

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Adit1122/conveyor-belt-recycling-detector"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

For non-causal architectures, replace AutoModelForCausalLM with the appropriate AutoModel class.