Caption: 

This image depicts a detailed scientific poster presented by the Toronto Computational Imaging Group (TCIG) under the University of Toronto and Vector Institute. The poster, titled "TurboSL: Dense, Accurate, and Fast Neural Inverse Structured Light," summarizes innovative research by Parsa Mirdehghan, Maxx Wu, Wenzheng Chen, and David B. Lindell.

The key objective of TurboSL is to reconstruct 3D surfaces efficiently from sequences of images captured under multiple Structured Light (SL) patterns. By leveraging high-frequency patterns and inverse rendering techniques, TurboSL aims to deliver fast, accurate, and dense 3D reconstructions. The poster outlines several components of their approach, including:

1. Pixelwise Disparity Estimation: This method minimizes the Zero-Mean Normalized Cross-Correlation (ZNCC) between projector and camera pixels to ensure high-quality surface reconstruction.
2. Alternatives for 3D from SL: The team compares 2D MLP and 3D MLP approaches for depth maps and density fields, highlighting the advantages and limitations of each.
3. TurboSL Neural Decoder: This section explains the decoding process involving forward and reverse rendering, and the impact of reflectance, cosine factors, and depth maps on reconstructions.
4. TurboSL Encoder: The encoder processes sample patterns, optimized for accuracy over precision, and incorporates a confusion matrix to refine the outputs.
5. Uncertainty Proxy by Binarization: This technique improves rendering by differentiating back-to-front and front-to-back processes to handle uncertainties effectively.

The poster includes visual aids and illustrative diagrams to explain their methods and results, demonstrating the capabilities and efficacy of TurboSL in advancing structured light techniques for 3D reconstruction.
Text transcribed from the image:
UNIVERSITY OF
TORONTO
TCIG V
Toronto Computational
Imaging Group
VECTOR
INSTITUTE
Key Objective: Fast and Accurate 3D from Multi-Pattern SL
o Reconstruct 3D surface from a sequence of images captured under N
Structured Light (SL) patterns while pursuing:
•
fast acquisition (only three or four SL images captured at 30-60Hz)
accurate surfaces (sub-pixel-accurate disparities)
TurboSL: Dense, Accurate a
Neural Inverse Structur
Parsa Mirdehghan, Maxx Wu, Wenzheng Chen, David B. Li
Our Approach: TurboSL by Neural Inverse Structured L
By recasting grayscale multi-pattern SL as an inverse rendering prob
or four projection patterns are enough for detailed geometric and pl
SL image 2
SL image 1
depth map
reflectance
dense reconstructions (mega-pixel-resolution depths)
Pixelwise Disparity Estimation
C
o Aims to minimize the Zero-Mean
Normalized Cross-Correlation (ZNCC)
between projector & camera pixels [1]
о Yields accurate surfaces given large N
•
requires long acquisition time to
capture many patterns
SL image 3
SL image 4
normal map
cosine factor
PERE
1. TurboSL Neural Decoder
forward rendering
scene
photo
cosine factor blurred pattern
reflectance
resi
N=30 [1]
N=3 [1]
t.odz
volume
rendering
+
reverse rendering
SL image
residual
reflectance cosine facto
camera projector
t
vo
re
o Quality degrades significantly with small N
Alternatives for 3D from SL: Scene Representation with MLP
2D MLP-Depth Map
3 patterns [1]
3D MLP - Density Field
3 patterns [1]
SDF
2. TurboSL Encoder
TurboSL works best with high-
frequency patterns optimized for
accuracy instead of precision [1]
sample pattern confusion matrix
3. Uncertainty Proxy by Bi
back-to-front
rendering
recon. r
fails to handle high-frequency
patterns
accurate image reconstruction,
but poor geometry
cam
proj
front-to-back
rendering