2.6 KiB
How it works?
0. Setup
pip install skylibs equilib
1. Data acquisition & pre-processing
OPTION 1: Real data
-
Acquire data using the apparatus (stick with two Ricoh cameras)
Make sure the cameras point up to avoid having lights in the seam.
Do two recordings with both Ricoh Theta cameras. Use the flash of a flashlight both times, the falling edge will later be detected by a software.
Cameras:
camera 1: left, well-exposed, serial: YN14100695
camera 3: right, under-exposed, serial: YN14111000 -
Take one video capture with the two cameras well-exposed.
-
Take one video capture with Camera 1 (left) well-exposed and Camera 2 (right) fast-exposed.
-
Run the script for automatic flash detection
It will temporally align the frames from the left and right camera and remove the initial frames with the flash.
It should output the following file structure
left_colmap_baseline/ 00000.png 00001.png ... 00XXX.png right_colmap_baseline/ 00000.png 00001.png ... 00XXX.png left_well_exposed/ 00000.png 00001.png ... 00XXX.png right_fast_exposed/ 00000.png 00001.png ... 00XXX.png
OPTION 2: Synthetic data
export PYTHONPATH=$pwd:$PYTHONPATH python lantern_scripts/split_synthetic_colmap.py --has_exposure --left_dir ./Dual_Cameras/left --right_dir ./Dual_Cameras/right --out_dir ./Dual_Cameras/split_colmap_out
2. Process data using COLMAP
ns-process-data lantern --data ./Dual_Cameras/split_colmap_subset --output-dir ./Dual_Cameras/split_colmap_subset/split_colmap_out5 --skip-image-processing --camera-type equirectangular --images-per-equirect 8
Thing to keep in mind: the way Nerfstudio splits train and test images may not be optimal: left might be in one dataset and right in another... that's a bit of a data leak.
2. Pre-training NeRF
Using both the well-exposed and fast-exposed frames, as well as their associated camera poses, a NeRF will be trained to predict both the RGB and mask to do volumetric rendering.
After convergence, go to the next step.
ns-train lantern-nerfacto --help
3. Running MomoNet
(Using which frames?) A pretrained UNet image-to-image network called MomoNet will infill the missing dynamic range and store these predictions in new images. It will be trained on the Laval indoor HDR dataset to do this task.
TODO: validate this part
4. Train from scratch or fine-tune the NeRF model using MomoNet's predictions