Skip to content

illum_est.md

Latest commit

 

History

History
68 lines (62 loc) · 2.18 KB

illum_est.md

File metadata and controls

68 lines (62 loc) · 2.18 KB

Illumination Estimation

Overview

The following scripts assume or create the following directory structure

|-- illum_est_expts
|   |-- data
|   |   |-- SamsungNX2000
|   |   |   |-- ours
|   |   |   |-- real
|   |   |   `-- upi
|   |   |-- Canon1DsMkIII
|   |   |-- Canon600D
|   |   |-- FujifilmXM1
|   |   |-- NikonD40
|   |   |-- NikonD5200
|   |   |-- OlympusEPL6
|   |   |-- PanasonicGX1
|   |   `-- SonyA57
|   `-- expts
|       |-- Canon1DsMkIII_illum_est_ours
|       |   |-- models
|       |   |-- tensorboard
|       |   `-- results
|   |-- nus_metadata
|   |   `-- nus_outdoor_gt_illum_mats
|   `-- synthia
|       `-- SYNTHIA_RAND_CVPR16
            `-- RGB

Prepare real data

For all methods, we use a subset of the SYNTHIA dataset for training and the NUS dataset for testing.

  • Prepare the NUS dataset
    • Follow the instructions in illum_est_nus.pptx for each camera and put the images under data/<camera>/real
    • From the NUS dataset webpage, download the groundtruth illuminant (MAT) files for each camera and put them under both data/nus_metadata/nus_outdoor_gt_illum_mats and data/<camera>/real for each camera
  • Download the SYNTHIA-RAND (CVPR16) dataset from link
    • We used 200 images from SYNTHIA_RAND_CVPR16/RGB for training and validation

Our method

Data generation

python3 -m jobs.generate_dataset_illum_est_graphics2raw

Training & Testing

python3 -m jobs.illum_est -c <camera1,camera2,...> -m ours

UPI

Data generation

Due to copyright issues, we cannot re-distribute third-party code. Please refer to upi.md before proceeding to the following steps.

python3 -m jobs.generate_dataset_illum_est_upi

Training & Testing

python3 -m jobs.illum_est_upi

Real

Data generation

Already completed in Prepare real data.

Training & Testing

python3 -m jobs.illum_est -c <camera1,camera2,...> -m real