Person Detection Demo

This repository demonstrates and compares person-segmentation approaches using local ONNX models.

Structure

• model_training/: Training scripts, ONNX export scripts, and utility scripts.
• models/: Exported model files (.onnx, .tflite).
• comparison/: Benchmark and video comparison script, plus generated output videos.
• video/: Input video used for inference.
• dataset/: Training, validation, and test data.

1. Model Comparison and Video Inference

Runs one or multiple ONNX segmentation models on video/demo_video.mp4, overlays masks, and writes output videos.

Run all models:

python comparison/comparison.py --model all --threads 1

Run a single model:

python comparison/comparison.py --model oneai --threads 1
python comparison/comparison.py --model unet --threads 1
python comparison/comparison.py --model deeplab --threads 1

Optional thresholds:

python comparison/comparison.py --model all --thr_oneai 0.5 --thr_unet 0.5 --thr_deeplab 0.5

Output:

• comparison/output_CustomCNN.mp4
• comparison/output_UNet.mp4
• comparison/output_DeepLabV3+.mp4

2. Theoretical Operations (FLOPs / Parameters)

Calculates theoretical Conv FLOPs and parameter counts from ONNX graph metadata.

python comparison/comparison.py --ops --model all

3. Training and ONNX Export

UNet training (person_demo.py)

python model_training/person_demo.py --epochs 100 --export_onnx

DeepLabV3+ MobileNet training (person_demo_deeplab.py)

python model_training/person_demo_deeplab.py --epochs 300 --export_onnx

4. Count ONNX Parameters

Counts total parameters from ONNX initializers.

python model_training/count_onnx_params.py

Requirements

• numpy
• opencv-python
• onnxruntime
• onnx
• torch
• torchvision
• pillow

Install:

pip install numpy opencv-python onnxruntime onnx torch torchvision pillow