This image features a scientific research poster titled "FSC: Few-Point Shape Completion," presented at the CVPR conference in Seattle, WA, held from June 18-22, 2023. The poster details a new model designed to complete point clouds using minimal data points. Key sections include:

1. **Contributions**: 
   - Investigation of completing point clouds with a few points.
   - Presentation of the FSC model for completing point clouds.
   - A dual-branch feature recovery architecture.
   - A two-stage revision module using WGAN.

2. **Few-Point Information Analysis**: 
   - An entropy analysis graph comparing input and completion results relative to the number of points.

3. **Pipeline**:
   - A detailed breakdown of the encoder, dual-branch, decoder, and feature extractor components in the model architecture.
   
4. **Experimental Results**: 
   - Tables showcasing the completion accuracy on various categories of ShapeNet and KITTI datasets. 
   - Methods compared include GRNet, PointR, PCN, SeedFormer, SVDFormer, and the proposed model.

5. **Visualizations**: 
   - Visual comparisons across different methods and input points on ShapeNet.
   - Figures illustrating the impact of removing branches in the feature extractor and varying input points.
   
The poster acknowledges contributions from the State Key Laboratory of Precision Measuring Technology and Instruments, Jianghan University, University at Buffalo, and Yangtze University, with authors listed as Xianzu Wu, Xiandong Wu, Tianyu Luan, Yajing Bai, Zhongyuan Lai, and Junsong Yuan. Notably, the corresponding author is Lai Zhiyu (laizhy@jhun.edu.cn). The poster board is labeled with the number 187.
Text transcribed from the image:
EXI
ANGHAN
Contributions
UNIVERSITY
YANGTZE
We investigate the potential of completing
point clouds with a few points.
UNIVERSITY
We present the FSC model capable of
completing a point cloud from a few points.
We design a dual-branch feature recovery
architecture to separately capture the
extensive and salient information.
We design a two-stage revision module that
Pipeline
Encoder
Input
x
Shared MLP
Shared MLP
FSC: Few-point Shape Completion
Xianzu Wu¹, 3*, Xianfeng Wu¹*, Tianyu Luan², Yajing Bail, Zhongyuan Lai¹*, Junsong Yuan²
State Key Laboratory of Precision Blasting, Jianghan University University at Buffalo, 3Yangtze University
Peint
feature
Extensive Branch
Point
feature
Feature
Global
feature
Global
feature
Revision
Point Cloud
Revision
Coarse output
YY
Polat
Salient Branch
feature
Global
feature
Experimental Results
feature
Global
feature
MLP + reshape
Coarse ground truth
EMD
Tile
Tile
CVPR
IN SEATTLE, WA JUNE 17-21, 2024
Deform
Tile
2D grid
Decoder
Detailed output Ground truth
Y
Visualizations
Inputs GRNet PoinTr PCN SeedFormer SVDFormer Ours Ground truth
Fig. 1 Results for 64 input points on the ShapeNet
w/o saliency branch w/o extensive branch
Dual branch
Ground truth
uses WGAN directly on the feature space
and on point cloud results.
Methods
GRNet
17.61
Table 1 64 points completion accuracy on the 8 seen categories of ShapeNet
Avg CD-1↓ Airplane Cabinet Car
19.40
23.32
Poin Tr
13.68
10.53
15.72
Few-Point Information Analysis
PCN
12.11
14.53
11.50
13.32
Chair Lamp
14.62 10.54
12.59 13.84
6.90
SeedFormer
12.38
8.56
16.14
Entropy (Shape)=
Number of Points
1
FPFH, log FPFH,
SVDFormer
Ours
11.22
7.21
7.89
3.47
100
Sofa Table
15.64 16.39 16.71
14.59 15.45
12.01 11.37
14.44
11.38 14.25 10.11 17.50 11.50 9.62
13.86 11.00 13.15 10.02 13.90 10.64 10.02
8.93
6.70 10.12 10.31 10.83 7.61 7.63
Vessel
24.26
14.701
11.98
12.80
100%
-Inputs
80
Completion Results
Methods:
GRNet
Entropy
60
70.3%
PCN
15.52
21.93
16.22
Table 2 64 points completion accuracy on the 8 unseen categories of ShapeNet
Avg CD-1↓ Bus Bed Bookshelf Bench Guitar Motorbike Skateboard Pistol
17.03
17.59 25.61
10.54 15.64
11.30 25.86
12.55 11.47
16.39
16.71
24.26
15.30
13.48
17.95
Poin Tr
14.92
14.62 20.44
16.84
40
14.22
11.56
13.40
14.45
13.80
45.5%
20
SVDFormer
Ours
13.60
13.30 21.47
13.73
12.90
7.94 22.60
12.98
11.69 8.00
9.59
13.25
9.91
17.44
GRNet
9.77
13.56
11.55
14.92
0
Original
2048
1024
512
Table 3 Completion results on KITTI dataset with various number of input points
Ours
6432
16
Methods
PCN
GRNet
Poin Tr
SeedFormer
AnchorFormer Snowflake AdaPoinTr
Ours
Number of Points
MMD
1.366
0.568
0.526
0.516
0.458
0.407
0.392
0.239
Fig. 2 Results of removing each branch in feature extractor
Ground truth
64
128
256
512
1024
2048
Fig. 3 Results of a lamp when the number of points changes
*Equal contribution. *Corresponding author (laizhy@jhun.edu.cn).
187