In the image, there is a large display of information exhibited at a convention or exhibition. The display features an array of charts, graphs, and diagrams, showcasing various aspects of a project or research study. The presentation appears to be discussing different forms of simulation and modeling, potentially related to manufacturing processes. The overall setup suggests that the convention is focused on advancing technologies and innovative solutions in the field of 3D modeling and simulational analysis. Attendees can gather around the display to learn about the project and its implications for the industry.
Text transcribed from the image:
CVPR
JUNE 17-21, 2024
wwww
SEATTLE, WA
Motivation
Highlight
UniMODE: Unified Monocular 3D Object Detection
Zhuoling Li¹
HKU
Xiaogang Xu2,3
2CUHK
Domain Head
SerNam Lim
3ZJU
Framework
Domain Confidence
(C₁₂C)
Hengshuang Zhao!
4UCF
Proposal Map
M Proposal Queries
Proposal Head
MLP
Extracted Feature
BXCXX
M+N Queries
N Random Queries
Feature Head
Indoor Scene
Outdoor Scene
Backbone + Neck
Sparse BEV Feature BEV Encoder
Projection
Self-Attn
Depth Head
Input Images
Bx3xHxW
Sparse BEV Feature Projection
Uneven BEV Feature Grid
BEV Decoderx6
Class Alignment Loss
Detection Results
HO
Query FFN
FFN
DALN Cross-AttnDALN
0-0
DALN
Domain Confidence.
(C₁.C₂C)
a
Sparse Tokens
Domain Parameters
(a₁. B₁)
(B)
Even Grid (Previous)
Visualization
VERSIT
ARKitScenes
Hypersim
nuScenes
191
Synthesized Scene
Annotation Problem
Although numerous 3D object detectors have been
developed, they are mostly designed for a single domain.
Unifying indoor and outdoor detection is challenging due
to diverse geometry distributions and heterogeneous
domain distributions.
Unstable Training
50
45
40
Loss boost
35
30
5
0
5
!(a.B) Input-dependent Parameter
Dense Feature
Sparse Feature Projection
on Uneven Grid
Input Layer Norm
Feature
Mini-adjust Feature
Output
Uneven Grid (Ours)
Images from Diverse Domains
In this work, we propose the UniMODE detector, which achieves SOTA in unified 3D object detection.
Geometry Distribution Gap
100
100
UniMODE
PETR
80
80
Gradient NaN
60
60
2
40
20
40
20
Heterogeneous Domain Distributions
Not labeled
KITTI
Objectron
SUN-RGBD
Main Results
Comparison with Existing Detectors
Method. AP AP AP AP AP AP AP AP
M3D-RPN 10.4%
SMOKE 19.5%
966
FCOS3D 17.6%
985
PGD
22.9%
11.29
GUPNet 19.9%
Im VoxelNet
21.5%
9.49
BEVFormer
PETR
25.9%
x
x
x
27.8%
X
x
X
I
Cube RCNN
9.5%
24.9%
15.0%
31.9%
27.9% 12.1%
8.5%
23.3%
UniMODE
UniMODE
22.3%
39.1%
28.3%
41.0%
26.9% 30.2%
7.4%
10.6%
29.7%
12.7%
8.1%
25.5%
8.7%
149%
31.1%
2825
5
0
2000
4000
6000
8000
10000
12000
-60
40
-20
O
20
40
-60
40
-20
20
Iteration
X axis
X axis
Outdoor
Although there are many popular 3D object
detectors, we find that they cannot converge
smoothly in unified 3D object detection.
By contrast, by incorporating our proposed
techniques, the developed detector UniMODE
achieves stable training.
Indoor
• Two-Stage Detection Architecture: Utilizing the first
stage network to inform the second stage about rough
target distribution, the training is stabilized.
• To address the grid size conflict between indoor and
outdoor scenes, we propose the uneven BEV grid.
(a) ARKitScenes
(b) Hypersim
Domain Adaptive Layer Normalization: An efficient
adaptive feature normalization method is developed
to bridge the significant feature discrepancy between
various data domains.
Class Alignment Loss: We devise an effective loss to
address the label conflict between various datasets.
Backbone AP AP AP AP AP 1 AP1
21.0%
DLA34
ConvNext 23.0%
6.7% 42.3% 52.5% 27.8% 31.7%
8.1% 48.0% 66.1% 29.2% 36.0%
Ablation Study
PH UBG SBFP UDA AP AP APsot Improvement
10.9% 14.3%
12.3%
13.4% 22.2%
15.9%
3.657
14.00% 23.8%
16.6%
0.7%
13.4% 23.7% 16.6%
0.0% t
14.8% 24.5% 17.4%
0.8%