Skip to content

VisionDepth3D v3.3 - Release

Choose a tag to compare

View all tags
@VisionDepth VisionDepth released this 12 Aug 13:42
· 71 commits to Main-Stable since this release
Release-v3.3
149e3af
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

VisionDepth3D v3.3 — Stability, Accuracy & Artifact Reduction

This update is a major overhaul to both the Depth Estimation Pipeline and the 3D Rendering Pipeline, with a focus on stability, accuracy, and artifact reduction.

How to Install

Because this package is larger than 2 GB, the installer is split into multiple files.

  1. Download VisionDepth3D_v3.3_WIN_x64_SETUP.exe
    and all accompanying .bin files from the release page
  2. Place them together in the same folder
  3. Run the .exe and follow the on-screen instructions

Depth Pipeline Updates

Black Bar Cropping for Depth Estimation

  • New ignore_letterbox_bars detects bars in the first non-empty frame
  • Crops top/bottom bars before sending frames to the depth model
  • Re-applies bars after processing with neutral depth values, preventing black regions from appearing closer or farther than the main scene

Output Resolution Preservation

  • Depth maps resized back to the original cropped resolution before re-adding bars
  • Ensures final depth video matches original width/height

Safety Checks

  • If bars exceed frame height or the frame is empty, bars reset to zero to prevent OpenCV errors

Unified Depth-to-Grayscale Conversion

  • convert_depth_to_grayscale() now handles:
    • PIL.Image
    • torch.Tensor
    • numpy.ndarray
  • Cleans NaN values and fixes shape inconsistencies
  • Centralized for consistent grayscale output

Sidecar Metadata for Bars

  • Saves .letterbox.json with top, bottom, and original_resolution next to the depth video

3D Pipeline Updates

Single or Batch Processing

  • Process one video or queue multiple for 3D rendering

Stability & Robustness

  • Render loop wrapped in try/except/finally for guaranteed cleanup
  • Defensive init for ffmpeg_proc and out
  • Early exit if VideoWriter fails
  • Pause handling keeps frame index + ETA/FPS updated
  • Cancel paths work during processing and pause
  • Automatic codec fallback if FFmpeg encoder is invalid

Depth Map Processing

  • TemporalDepthFilter (EMA smoothing) reduces depth flicker
  • Percentile-based normalization for consistent depth range
  • Midtone shaping (gamma) improves depth layering
  • Optional curvature enhancement for roundness

Stereo / Parallax Control

  • ShiftSmoother damps rapid disparity changes
  • Edge-aware masking + feathering reduces tearing
  • Dynamic IPD scaling adapts stereo strength
  • Subject-tracked zero parallax with window easing
  • Optional dynamic convergence bias
  • IPD factor knob for global stereo strength

Image Quality

  • GPU depth-of-field blur with Gaussian blending
  • Brightness-preserving sharpening with highlight protection

Framing, Aspect & Output Formats

  • Aspect-ratio safe resizing with pad_to_aspect_ratio
  • Two modes:
    • Preserve Original Aspect Ratio
    • Target Output Aspect
  • Formats:
    • Full-SBS
    • Half-SBS
    • VR 1440×1600
    • Dubois anaglyph
    • Passive interlaced

Encoding & I/O

  • FFmpeg over stdin:
    • CRF for libx*
    • CQ for NVENC with -b:v 0
  • CPU/GPU encoder mapping, OpenCV fallback

UX / Telemetry

  • Smooth, real-time progress/FPS/ETA — also while paused
  • More descriptive logging

Download

📥 Download VisionDepth3D v3.3

VisionDepth3D is free for personal and non-commercial use.
Commercial use or redistribution without consent is prohibited.

What's Changed

Full Changelog: Release-v3.2.4...Release-v3.3

Contributors

VisionDepth
Assets 5
Loading