Food Instance Segmentation

NCKU "Introduction to Neural Networks" – Final Project

---

---

Categories

1. Staple Food
2. Main Course
3. Side Dish
4. Vegetable

---

Dataset Description

Download Dataset

Category	Instances
Staple Food	1174
Main Course	1022
Side Dish	965
Vegetable	1268
Total	4429

---

Labeling Tools

• Detection: LabelImg
• Segmentation: LabelMe

---

Dependencies & Setup

Environment: pytorch-21.06-py3

1. Download Pretrained Weights

wget -q https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_C4_3x/137849393/model_final_f97cb7.pkl 

2. Install Dependencies

sudo apt-get update
sudo apt-get install ffmpeg libsm6 libxext6 -y

---

Code Overview

show_images.py

Check dataset images and annotations

python show_images.py

---

train.py

Train and evaluate the segmentation model

python train.py

---

inference.py

Visualize predictions on test images

python inference.py

---

Experiment Results

1. BBox Evaluation

AP	AP50	AP75	APs	APm	APl
61.920	80.313	74.643	12.612	47.781	68.195

2. Per-Category BBox AP

Category	AP
Staple Food	78.647
Main Course	52.682
Side Dish	48.953
Vegetable	67.399

---

3. Segmentation Evaluation

AP	AP50	AP75	APs	APm	APl
44.442	78.253	49.255	2.878	27.602	51.990

4. Per-Category Segmentation AP

Category	AP
Staple Food	57.265
Main Course	34.415
Side Dish	32.760
Vegetable	53.329

---