Skip to content

Latest commit

 

History

History
146 lines (110 loc) · 4.28 KB

Readme.md

File metadata and controls

146 lines (110 loc) · 4.28 KB
Raw

Nutrition: Coarse-to-Fine Nutrition Prediction

This repository is the official implementation of our method. In this work, we propose the coarse-to-fine paradigm combined with structure loss to predict nutrition.

Requirements

To install requirements:

conda env create -f environment.yaml

Before running the code, please activate this conda environment.

Data Preparation

Download food images of Nutrition5K dataset and our best model from Baidu Netdisk.

acess code：gw8x

Please ensure the data structure is as below

Nutrition5K,ECUSTFD,VFD
├── data
   └── Nutrition5K
       └── imagery
           └── new_realsense_overhead           
               ├── dish_1556572657
                   └── rgb.png
               ├── dish_1556573514                  
                   └── rgb.png
               └── ...
   └── ECUSTFD
       ├── TrainImage           
           ├── apple001S(1).JPG
           ├── apple001S(2).JPG
           └── ...
       └── TestImage
           ├── apple015S(1).JPG
           ├── apple015S(2).JPG
           └── ...
   └── VFD
       ├── VFDS-15
           ├── train_resize
               ├── 1.jpg
               ├── 3.jpg
               └── ...
           ├── test_resize
               ├── 2.jpg
               ├── 9.jpg
               └── ...
       └── VFDL-15
           ├── train_resize
               ├── 1.jpg
               ├── 3.jpg
               └── ...
           ├── test_resize
               ├── 2.jpg
               ├── 9.jpg
               └── ...

Note: the directory 'new_realsence_overhead' is generated by the python script 'Nutrition5K/utils/resize_img.py' before training model. The same is true for the ECUSTFD and VFD datasets.

Training & Evaluation

To train the our model on Nutrition5K dataset, please run this command:

cd .projects/Nutrition5K/Nutrition5K
python main.py

To train the our model on ECUSTFD dataset, please run this command:

cd .projects/ECUSTFD/code
python main.py

To train the our model on VFD dataset, please run this command:

cd .projects/VFD/code
python main.py

Evaluation

To evaluate our model on Nutrition5K, run:

cd .projects/Nutrition5K/Nutrition5K/criterion
python compute_eval_statistics.py

To evaluate SBF-Net model on ECUSTFD, run:

cd .projects/ECUSTFD/code/criterion
python compute_eval_statistics.py

To evaluate SBF-Net model on VFD, run:

cd .projects/VFD/code/criterion
python compute_eval_statistics.py

Pre-trained Models

You can download pretrained models and our datasets here: Baidu Netdisk.

Results

Our model achieves the following performance:

Nutrition5K

Nutrition Calories Mass Fat Carb Protein
pMAE 25.0 19.7 37.4 32.3 34.5

ECUSTFD

Food Bread Sachima Fried_wist Litchi Mooncake Plum Qiwi Egg Bun Mango
pMAE 16.7 6.9 10.1 7.2 18.3 7.7 11.3 8.3 13.0 5.3
Food Lemon Peach Doughnut Banana Orange Tomato Pear Grape Apple Mean
pMAE 0.5 15.9 8.7 15.4 5.7 6.2 2.5 12.5 2.7 9.2

VFD

VFDS-15

class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Mean
pMAE 12.8 7.4 10.6 9.0 9.4 8.6 8.3 8.1 7.3 7.0 6.5 5.8 6.0 4.9 3.7 6.4

VFDL-15

class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Mean
pMAE 12.3 10.9 10.9 10.0 7.8 9.8 9.0 6.2 6.0 6.0 4.6 5.8 6.0 5.3 7.2 6.9