Skip to content

Latest commit

 

History

History
405 lines (242 loc) · 15.6 KB

README.md

File metadata and controls

405 lines (242 loc) · 15.6 KB
Raw

iw3

I want to watch any 2D video as 3D video on my VR device, so I developed this very personal tool.

iw3 provides the ability to convert any 2D image/video into side-by-side 3D image/video.

This project is under construction.

日本語の説明

Overview

Usage

First see the installation section.

GUI

Run iw3.gui module from the root directory of nunif.

python -m iw3.gui

On Windows, run Run iw3 GUI.bat.

iw3-gui

CLI

Run iw3 or iw3.cli module from the root directory of nunif.

python -m iw3 -i <input file or directory> -o <output file or directory>

The following command shows all available options.

python -m iw3 -h

Also, when running iw3 for the first time, it may take some time as it needs to download large model files.

What is --divergence(-d) option?

(3D Strength in GUI)

divergence

2.0 by default.

This parameter changes the best distance for the screen position.

With lower value, the best distance for the screen position is farther away. With higher value, the best distance for the screen position is closer. However, with higher value, artifacts are more visible.

You can adjust the screen position by zoom-in/zoom-out on the VR Video Player.

What is --convergence(-c) option?

convergence

  • 0 is good, but screen edge areas are hard to see.
  • 1 is the most friendly for curved display setting.
  • 0.5 by default.

What is --ipd-offset option?

(Your Own Size in GUI)

ipd-offset

This may be adjustable on the VR Player. If so, set it to 0 (by default).

What is --foreground-scale option?

When specifying a positive value (1 .. 3), foreground depth is scaled up and background depth is scaled down.

foreground-scale

To be used for outdoor photos where foreground(people) look very flat. For videos, 0(by default) is recommended.

When specifying a nagative value (-1 .. -3), background depth is scaled up and foreground depth is scaled down.

foreground-scale-negative

Note that the transformation formula is different for ZoeDepth models(ZoeD_N, ZoeD_Any_N) and DepthAnything models(Any_S, Any_B, Any_L), even for the same value.

What is --edge-dilation option?

(Edge Fix in GUI)

This parameter is used only for DepthAnything models (Any_S, Any_B, ANY_L).

DepthAnything model outputs very accurate depth, but in stereo generation, it causes artifacts at foreground and background edges.

This approach reduces artifacts by dilating foreground segments (high value area).

edge-dilation

0 is disabled. 2 by default. 4 is the most eye-friendly, but it degrades depth accuracy.

About Colorspace

See Colorspace and nagadomi#164 .

About Depth Resolution

See nagadomi#168 .

About Video Format

mp4 is a highly compatible format.

mkv can be previewed during conversion.

avi is provided to output lossless video for your own encoding.

Video Codec

libx264 is for H.264 which is a highly compatible format. However, at higher resolutions like 4K, the file size will be larger, which may cause playback lag/artifact.

libx265 is for H.265.

utvideo is for lossless video. You may need Ut Video Codec Suite for playback.

Level option for H.265

auto is recommended. This will cause an error if the wrong level is selected for a video that is out of range for the specification.

H.264 Profile

There is no way to specify H.264 profile. However, it seems to be Constrained Baseline when --preset ultrafast and High otherwise.

About VR Player

I have tested the results with the following software.

Pigasus VR Media Player

Pigasus works perfectly for SBS 3D videos, images, and Samba drive(SMB). However, I am not a big fan of its user interface and user experience.

If you can only choose one software, I would recommend this one.

SKYBOX VR Video Player

With recent updates(v1.1.6), most features of 3D Full SBS are now working. However, the following features have not yet been implemented.

  • No ability to navigate prev/next images with joystick
  • Screen height position is not adjustable

About file naming rule

VR Player detects media format by filename. Adding _LRF_Full_SBS suffix to the filename will identify the file as full side-by-side 3D media.

When specifying a directory with -o option, it is automatically output as a filename with {original_filename}_LRF_Full_SBS.(png|mp4).

Reference:

I confirmed that _LRF_Full_SBS works with all of the above software.

VR180 format

When --vr180 option is specified, the video is output in VR180 format (equirectangular).

This is usually not recommended because of poor usability during playback.

This is useful if your video player does not have the ability to play Full SBS 3D videos or if you want to post the video on Youtube.

Half SBS format

When --half-sbs option is specified, the video is output in Half SBS format (subsampled at half resolution).

Older VR devices may only support this format. Also, you may need to add _3dh_ to the filename to play it.

Full TB and Half TB format

When --tb or --half-tb option is specified, the video is output in TopBottom format.

TopBottom format can be played back with higher resolution than SBS on some 3D TVs (Polarized/Passive 3D system).

Anaglyph 3D format

When --anaglyph option is specified, the video is output in Red-Cyan Anaglyph 3D format.

(On GUI, select Anaglyph * option in Stereo Format)

Currently, the following methods are supported.

Method
color Color anaglyphs. Partial color reproduction. Retinal rivalry.
half-color Half-Color Anaglyphs. Partial color reproduction. Less retinal rivalry than color anaglyphs.
gray Grayscale Anaglyphs. No color reproduction.
wimmer Wimmer's Optimized Anaglyphs.
wimmer2 Wimmer's Improved Method.
dubois Dubois Method.
dubois2 Dubois Method. This version ignores the specification The first and second terms should be clipped to [0,1] before adding.

Reference:

For video, I recommend --pix-fmt yuv444p or --pix-fmt rgb24 option. yuv420p (by default) uses 4:2:0 chroma subsampling, red colors are degraded and cause ghosting artifacts (crosstalk).

JPEG have the same problem so I recommend using PNG (by default) instead.

Also, --convergence 0.5 --divergence 2.0 is recommended.

Export and Export Disparity

Export and Export Disparity are features to output depth and frame images.

See nagadomi#97 (comment) for details at the moment.

Trouble shooting

Very flat foreground

This tends to happen with outdoor scene photos.

There are several ways to fight this problem.

  • Try --foreground-scale 3 option
  • Try--remove-bg option
  • Try combined option --divergence 4 --convergence 0 --foreground-scale 3 --remove-bg

When --remove-bg is specified, the background area is removed using rembg with U2-net's human segmentation model, before estimating depthmap.

Video encoding error

Please post to the issue about the format of the video.

Large video file size

You can reduce the file size with --preset medium option. --video-codec libx265 also helps reduce the file size.

60fps video drops to 30fps

By default, FPS is limited to 30fps. Use --max-fps 128 option.

Note that 60fps video takes twice as long to process as 30fps limit.

Problems with older GPUs

FP16 can cause slowdowns and errors on older GPUs (older than GeForce 20 series).

In CLI, FP16 can be disabled with --disable-amp option.

It's a giant!

This is a problem with SBS 3D video that it cannot be rendered in actual size scale.

You can try adjusting scale manually.

  • Adjust IPD offset on VR Player
  • Use --ipd-offset option(You own size in GUI) to adjust IPD offset

It is better to adjust IPD offset on VR Player, but you can also apply IPD offset to the output image.

On SKYBOX Player, set the 3D effect slider to around < -0.3.

On Pigasus, set Settings > Advanced Settings > IPD slider to large.

Also, on Pigasus, you can zoom and pan the image by double-clicking the trigger button on the VR controller.

CUDA Out of Memory

Use --low-vram option.

I tested this program on RTX 3070 Ti (8GB VRAM, Linux) and GTX 1050 Ti (4GB VRAM, Laptop, Windows). Both work with the default option.

How to convert rotated(height width swapped) video correctly

Use --rotate-left(rotate 90 degrees to counterclockwise) or --rotate-right(rotate 90 degrees to clockwise) option to fix the rotation.

Video is interlaced

Use --vf yadif option to deinterlace the video source. You can use ffmpeg's video filter with --vf option.

Note that

  • only serial pipeline is supported

See https://ffmpeg.org/ffmpeg-filters.html

Artifacts in encoded video

When preset=slower|veryslow|placebo is used for high resolution video, it may produce video files that are not supported by hardware decoder. If you have been using that setting, try preset=medium.

Recommended workflow for slow processor or very large video

First check the results with a few samples. There are two ways.

The following command processes video keyframes about every 4 seconds and outputs the result as 3D photos (image files).

python -m iw3 --keyframe --keyframe-interval 4 -i input_video.mp4 -o output_dir/

The following command processes video frames every 2 seconds and outputs the result as 3D slideshow-like video (video file).

python -m iw3 --max-fps 0.5 -i input_video.mp4 -o output_dir/

If the results are acceptable, process the full video.

Limitation

--method row_flow(by default) is currently only trained for the range 0.0 <= divergence <= 2.5 and 0.0 <= convergence <= 1.0.

About row_flow model and its training

See nagadomi#60 .

Basically, fine tuning for this model is not necessary. Perhaps what is needed is fine tuning for ZoeDepth.

Monocular Depth Estimation Models

Short Name
ZoeD_N ZoeDepth model NYUv2. Tuned for indoor scenes.
ZoeD_K ZoeDepth model KITTI. Tuned for outdoor scenes (dashboard camera view).
ZoeD_NK ZoeDepth model NYUv2 and KITTI.
ZoeD_Any_N Depth-Anything Metric Depth model NYUv2. Tuned for indoor scenes.
ZoeD_Any_K Depth-Anything Metric Depth model KITTI. Tuned for outdoor scenes (dashboard camera view).
Any_S Depth-Anything model small. The most efficient model, with promising performance.
Any_B Depth-Anything model base.
Any_L Depth-Anything model large. This model gives high quality, but is also heavy in terms of computation.
Any_V2_S Depth-Anything-V2 model small.
Any_V2_B Depth-Anything-V2 model base. (cc-by-nc-4.0)
Any_V2_L Depth-Anything-V2 model large. (cc-by-nc-4.0)
Any_V2_N_S Depth-Anything-V2 Metric Depth model Hypersim small. Tuned for indoor scenes.
Any_V2_N_B Depth-Anything-V2 Metric Depth model Hypersim base. Tuned for indoor scenes
Any_V2_N_L Depth-Anything-V2 Metric Depth model Hypersim large. Tuned for indoor scenes. (cc-by-nc-4.0)
Any_V2_K_S Depth-Anything-V2 Metric Depth model VKITTI small. Tuned for outdoor scenes (dashboard camera view).
Any_V2_K_B Depth-Anything-V2 Metric Depth model VKITTI base. Tuned for outdoor scenes (dashboard camera view).
Any_V2_K_L Depth-Anything-V2 Metric Depth model VKITTI large. Tuned for outdoor scenes (dashboard camera view). (cc-by-nc-4.0)

Personally, I recommend ZoeD_N, Any_B or ZoeD_Any_N. ZoeD_Any_N looks the best for 3D scene. The DepthAnything models have more accurate foreground and background segmentation, but the foreground looks slightly flat.

For art/anime, DepthAnything is better than ZoeDepth.

About Any_V2_B ,Any_V2_L, Any_V2_N_L, Any_V2_K_L

These models are licensed under cc-by-nc-4.0 (Non Commercial). These are not available by default due to conflicts with nunif MIT license.

If you want to use it, agree to the pre-trained model license and place the checkpoint file yourself.

Short Name Path
Any_V2_B iw3/pretrained_models/hub/checkpoints/depth_anything_v2_vitb.pth
Any_V2_L iw3/pretrained_models/hub/checkpoints/depth_anything_v2_vitl.pth
Any_V2_N_L iw3/pretrained_models/hub/checkpoints/depth_anything_v2_metric_hypersim_vitl.pth
Any_V2_K_L iw3/pretrained_models/hub/checkpoints/depth_anything_v2_metric_vkitti_vitl.pth

These files can be downloaded from Models section of https://huggingface.co/depth-anything .

Stereo Generation Method (Left-Right Image Generation)

Short Name
row_flow_v3 Calculating the backward warping parameters with ML model. Trained with 0.0 <= divergence <= 5.0 and synthetic training data generated by forward_fill. Default method.
row_flow_v2 Previous default model. Trained with 0.0 <= divergence <= 2.5 and synthetic training data generated by stable-diffusion-webui-depthmap-script 's method.
forward_fill Depth order bilinear forward warping. Non ML method.
row_flow_v3_sym Calculating the backward warping(grid_sample) parameters with ML model. The left and right parameters are fully symmetric. 2x faster than row_flow_v3. For experimental use.
forward forward_fill without hole fill. For experimental use.
grid_sample,backward Naive backward warping. Lots of ghost artifacts. For experimental use.

Updating torch.hub modules (ZoeDepth, DepthAnything)

When you manually update the code, run the following command.

python -m iw3.download_models

The command syncs the following repositories.

If you already downloaded the model files (checkpoint filess), downloading model files will be skipped.