This repository contains the code and data for the proposed SPADE-SR (Spatial Adaptive DEnormalization with Self-Sampling) model and LRT-Human (Low-Resolution Thermal Human) dataset for semantic image synthesis using low-resolution thermal heatmaps from a thermopile array sensor.

Download and unzip the dataset. [Google Drive]

rgb: RGB images

thermal: thermal heatmaps

thermal_calibrated: calibrated thermal heatmaps (unused)

thermal_noise05: thermal heatmaps with noise (SD=0.5 Celsius) added (unused)

thermal_noise10: thermal heatmaps with noise (SD=1.0 Celsius) added (unused)

The LRT-Human dataset consists of ~22k low-res thermal heatmap and RGB image pairs of a single human under various scene configurations in a way that thermal and non-thermal attributes can be spotted explicitly. Each folder in the dataset contains data in a single scene, presenting the same non-thermal attributes. The .npy array file is available for faster data reading. However, it is shuffled and in the shape of [B,W,H,C], which means it is not grouped in folders.

The codes are tested in Python version 3.6.7 and Tensorflow==1.15. Please run

pip install -r requirements.txt

to install packages.

To train a model, use the code train.py with a configuration file. Please find the folder cfgs for example config files. The following code would train a standard SPADE-SR model:

python train.py cfgs/SPADE_SR_64.yaml

You could fill in .h5 weight files to the model fields in the config file to continue a training as well.

If you want to calculate the FID score on the fly whenever the model is saved, in the config file, set CALC_FID: true and REAL_STATS_PATH to your .npz feature statistics file. To prepare the pre-calculated feature statistics for your dataset's real images, edit and run the code precalc_stats_example.py. Please be careful when processing the image data using OpenCV (cv2), which reads and writes an image in the BGR channel order. The LRT-Human dataset's .npy array file is in BGR channel order, to reverse the order, simply index [:,:,:,::-1], which reverses the order at the channel axis.

To train an inversion encoder after the GAN training, use the code inversion.py with a configuration file. Please find the file cfgs/inversion.yaml for available fields. There are two options to set the config file:

(Option 1) The Standard Method:

python inversion.py your-inversion-config.yaml

(Option 2) The Auto-Parsing Method: It's common that many fields in the config file are copies between the GAN and inversion training phase. This method would conveniently auto-parse the fields whose values are commonly the same directly from a GAN training config. Just provide a GAN training folder after the inversion config:

python inversion.py cfgs/inversion.yaml your-training-folder

A field whose value is set to None in the inversion config would be replaced by the field in the GAN training config according to the field relation list cfg_keypair_list in inversion.py. The list covers commonly used fields and rarely needs to be edited. You have to manually set the GENERATOR and DISCRIMINATOR fields though (just the .h5 filename itself without path, because the rest would be auto-filled for you).

To generate the k times n evaluation grid, use the code evaluate.py with a configuration file.

python evaluate.py your-evaluation-config.yaml

Please find the file cfgs/evaluate.yaml for available fields.

You can set the config using auto-parsing as well:

python evaluate.py cfgs/evaluate.yaml your-INVERSION-folder

To calculate the reconstruction error according to the paper, use:

python calc_mae.py your-calc-config.yaml

Please visit cfgs/calc_mae.yaml for example config.

To visualize real and fake image samples, use the code generate_image_samples.py with a configuration file:

python generate_image_samples.py your-gen-config.yaml

Please visit cfgs/generate_image_samples.yaml for example config.

You can set the config using auto-parsing as well:

python generate_image_samples.py cfgs/generate_image_samples.yaml your-INVERSION-folder

The C code for the Arduino sensor MCU is at grideye_read/grideye_read.ino. Connect the MCU to the computer and use data_processing/collect.py to collect RGB images and temperatures. Use data_processing/render.py to transform temperature values into heatmaps and perform the alignment. data_processing/align.py provides a simple combination-based method for optimal alignment parameter search utilizing additional human segmentation masks from MASK-RCNN. These codes are just for reference and need cleaning-up and better documentation. It's recommended to write your own data collection pipeline.

FID codes are modified from TTUR

Arduino MCU code are modified from SparkFun