Joint Optimization of Camera Model and Deep Neural Network for Image Recognition (WACV 2026)

Official implementation for HFD100 hyperspectral image classification with joint optimization of camera model and a classification network.

Paper: Joint Optimization of Camera Model and Deep Neural Network for Image Recognition (WACV 2026)

Overview

• Train and evaluate only (train.py + evaluate.py)
• Supports HFD100 Flower and Leaves datasets
• Preprocessing required: Download the original dataset, convert to HDF5, then compute sRGB max values (see below)

Setup

poetry install

Dataset and Preprocessing

Preprocessing must be run first before training and evaluation.

1. Download Original Dataset

Download the HFD100 dataset (Zheng, Zhang & Fu, Knowledge-Based Systems, 2022). The dataset (~55GB, .mat format) is hosted on Baidu Cloud.

2. Convert to HDF5

Convert .mat files to HDF5 format to produce:

• MatFlower60.h5
• MatLeaves60.h5
• (Optional) MatScenes60.h5

Use the provided conversion script:

python data/hfd100tohdf5.py \
  --input-root /path/to/HFD100_Mat_dataset \
  --output-dir ./data \
  --datasets flower leaves

The generated HDF5 files have the following structure:

train/
  metadata      (JSON)
  hs/
    image0001, image0002, ...
test/
  metadata
  hs/
    ...

3. Compute sRGB Max Values

export HFD100_DATA_DIR=/path/to/your/hdf5/directory

python data/compute_srgb_max.py HFD100_Flower -d $HFD100_DATA_DIR -o data/HFD100_Flower_srgb_max_values.json
python data/compute_srgb_max.py HFD100_Leaves -d $HFD100_DATA_DIR -o data/HFD100_Leaves_srgb_max_values.json

Output JSON files can be placed in either $HFD100_DATA_DIR or data/ in the project root.

Quick Start

Training

export HFD100_DATA_DIR=/path/to/your/hdf5/directory

# Flower
python src/train.py --config src/configs/flower.yaml

# Leaves
python src/train.py --config src/configs/leaves.yaml

Evaluation

1. Edit src/configs/eval_config.yaml:

run_parameters:
  train_run_dir: runs/train/your_run_name
  checkpoint: best_model  # or latest
  output_dir: flower_eval

2. Run:

python src/evaluate.py --config src/configs/flower.yaml --eval_config src/configs/eval_config.yaml

Pretrained Weights

Classification backbones (ResNet-18, ViT-S/16, SE-ResNet50) pretrained on HFD100 Flower/Leaves are hosted on Hugging Face and downloaded automatically based on classification_model and dataset_name.

Main Files

File	Purpose
src/train.py	Training
src/evaluate.py	Evaluation
src/hsi_dataset.py	HFD100 dataset loader
data/compute_srgb_max.py	Preprocessing: sRGB max value computation
src/configs/flower.yaml	Flower config
src/configs/leaves.yaml	Leaves config

Repository Layout

src/
├── train.py
├── evaluate.py
├── hsi_dataset.py
├── configs/
│   ├── flower.yaml
│   ├── leaves.yaml
│   └── eval_config.yaml
├── models/
├── utils/
└── camera_parameters/
data/
├── compute_srgb_max.py
├── HFD100_Flower_srgb_max_values.json  (generated)
└── HFD100_Leaves_srgb_max_values.json  (generated)

Config / Environment

Config / Arg	Description
data_dir	HDF5 directory. Config or --data_dir. Default: ./data (or HFD100_DATA_DIR)
srgb_max_dir	Directory for sRGB max JSON files. Config or --srgb_max_dir. Default: same as data_dir

Citation

@InProceedings{Noboru_2026_WACV,
    author    = {Noboru, Youta and Ozasa, Yuko and Tanaka, Masayuki},
    title     = {Joint Optimization of Camera Model and Deep Neural Network for Image Recognition},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {March},
    year      = {2026},
    pages     = {7626-7635}
}