This image shows an academic poster presentation detailing the research on "GenNBV: Generalizable Next-Best-View Policy for Active 3D Reconstruction." The poster is displayed at a conference, indicated by the booths numbered 173 and 174 visible on either side. The presenting institution is The Chinese University of Hong Kong.

The central theme of the research is to develop and investigate the GenNBV policy, which is designed to improve the efficiency and effectiveness of 3D reconstruction processes. Key sections of the poster include:

- **Motivation:** Comparing existing NBV (Next-Best-View) policies and highlighting their limitations, such as limited action space and heuristic criteria, against the proposed GenNBV which offers free-action space and state embedding for cross-dataset generalization.
- **Overview:** An introduction to GenNBV, stating it as the first RL (Reinforcement Learning)-based end-to-end NBV policy allowing for free-space exploration.
- **Methodology:** Detailing the pipeline from state embedding to NBV policy representation, including probabilistic occupancy grid and a reward function focused on coverage ratio.
- **Generalizability Evaluation:** Illustrating how GenNBV is trained using the Houses3K training set and its application in varied environments.
- **Visualization Comparison and Ablation Study:** Presenting visual results comparisons and performance metrics, showcasing the superior results of GenNBV over existing policies.

The bottom of the poster includes QR codes linking to the project paper and its online page for further reading. Contact information for correspondence with the research team is also provided. The setting, with other posters and participants visible, suggests a lively and engaging academic conference atmosphere.
Text transcribed from the image:
PR
21, 2024
E, WA
173
上海人工智能实验室
Shanghai Artificial Intelligence Laboratory
香港中文大學
The Chinese University of Hong Kong
Motivation
Existing NBV policies
Limited action space (≤ 3 DoF)
No need to consider collision
Per-scene optimized representations
Indirect heuristic criteria
Small objects
NeR
"
GenNBV: Generalizable Next-Best-View Policy for Active 3D Reconstruction
Xiao Chen 1.2 Quanyi Li¹ Tai Wang Tianfan Xue2t Jiangmiao Pang
GenNBV (Ours)
1 OpenRobotLab, Shanghai Al Laboratory
Free action space (5 DoF)
Collision avoidance
Methodology
Haterical Observations
Generalizable state embeddings
"
Evaluation metrics as reward
Large objects and scenes
AFLAME
2 MMLab, The Chinese University of Hong Kong
Visualization Comparison
@@
Policy
An a
Sconrage Ralle
GenNBV
(Ours)
Reward Calculation
(a) Houses
CVPR
SEATTLE, WA JUNE 17-21, 2024
IN Onnige
GenNBV captures more self-occlusion details of cross-dataset
unseen objects and thus build high-quality meshes
174
Object-Centric Capturing
Train on Hundreds of Objects
Per-scene Training
Free-space Capturing Cress-dataset Generalization on Novel Objects
Our Generalizable Free-space NBV Policy
Previous NBV Policies
Overview
GenNBV is the first RL-based end-to-end NBV policy, which allows
free-space exploration and cross-dataset generalization.
GenNBV is guided by informative and generalizable state embedding.
GenNBV extend previous limited 3 DoF action to 5 DoF free space, that allow
adapt to any geometry and capture self-occlusion details.
GenNBV can be generalized cross-dataset and cross-category w/o finetuning.
Project Page: gennbv.github.io
Contact: cx123@ie.cuhk.edu.hk¦¦
Next-Best-View Capturing
Hemisphere
(CR: 89.8%)
Scan-RL
(CR: 90.6%)
GenNBV
(CR: 96.8%)
Pipeline:
extract state embedding from observations; 2 NBV policy predicts
next best viewpoint; 3 agent captures novel observations; Repeat 1-3.
■ Probabilistic occupancy grid indicates the occupancy and exploration process.
Depth-based Representation Mean AUC Final Coverage Ratio
Geometrie 2D Map
74885
Binary 30 Grid
77.29%
10.35%
Probabilistic 3D Grid
84565
Occupied
Free
Unknown
The effectiveness of probabilistic 3D grid
Reward function: Coverage Ratio (CR),
Binary 3D Grid Probabilistic 3D Grid
CR+1-CR, where CR-
Uncertainty-guided
Scan-PL.
Cumar Card
100%
GenNBV generalizes well f
o unseen challenging 3D scenes.
Ablation Study
Generalizability Evaluation
■GenNBV is trained on hundreds of houses from Houses3K training set
BLIPS
BLUS
FLIPS
SH
Nars
Category
NBY Puky AUCT FORT M
GIT
8.16
-
37 806
Tracks
Описука30
NS Bes
43
4
Scan-RL.
70% M
LIS
Dan OCD
140
BLIT
LA
AN
BLS MUS 431
11309
A
L
79
AM
G.M
MAYS
stare
834
Room
Replic
SUS
Scan-RL.
GNY BOS RS
1.12
"
199
1631
Cross-dataset + Cross-category
Takeaways:
Cross-dataset
1) Generalizability of NBV policies: RL-based > Info Gain-based > Heuristic
2) Key elements of RL-based policies: free action space and informative scene repre.
3) Further generalization: cross-dataset cross-category indoor/outdoor scenes
3119
Takeaways:
Geometry and semantics are
both significant for NBV policies;
The proposed probabilistic 3D
grid benefit the generalization;
Takeaways:
Increasing the diversity of
training objects improves the
generalizability of NBV policies.
Paper
Project Page
0.00