A detailed caption for this image could be:

A poster presentation titled "Neural Video Compression with Feature Modulation," authored by Jiahao Bu, Li Li, Bin Li, and Yan Lu from Microsoft Research Asia, displayed at the CVPR conference in Seattle, WA, June 17-21, 2024. The poster is organized into three main sections: Introduction, Methods, and Results.

**Introduction:** Features a graph comparing different video codecs, highlighting the superior performance of the Neural Video Codec (NVC) over ECM in long prediction chains. Key points include that NVC supports a large quality range for both YUV and RGB content, achieves 29.7% bitrate saving, and incurs a 16% MAC reduction over previous state-of-the-art NVC. 

**Methods:** The core framework of the conditional coding-based structure is illustrated, showing the process flow from the input frame to the frame coding function and temporal feature modulation stages. Diagrams detail the components and operational steps in the video compression method.

**Results:** Includes rate control graphics and a table comparing BD-rate percentages across various codecs, where lower values indicate better performance. Visual examples of video quality under different codecs are provided, comparing the original to ECM-5.0, DCVC, and DCVC-FM codecs. The GitHub link for the project is also provided: https://github.com/microsoft/DCVC.

The poster is neatly structured and visually communicates the advancements in neural video compression achieved by the research team.
Text transcribed from the image:
11.3 dB
PSNR (dB)
45
43
Introduction
DCVC-FM
41
DCVC-DC
3.6 dB
37
Bate
Decrease 20%
40%
DCVC-OC
HM-16.25
0%
0.004
80%
84.4%
42.8 %
ECM-11.0
VIM 170
HM-16.25
30.2%
17.2%
VTM-17.0 ECM-11.0 DCVC-FM
RPP
0.012
0.016
0.020
The first neural video codec (NVC) that surpasses
ECM under long prediction chain
Support a large quality range for both YUV and
RGB contents in single model
29.7% bitrate saving and 16% MAC reduction over
previous SOTA NVC
https://github.com/microsoft/DCVC
Neural Video Compression with Feature Modulation
Jiahao Li Bin Li Yan Lu
Microsoft Research Asia
X-2
Methods
t-1
Ct-2
F₂
fframe
ITcontext
fframe frcontext fframe
+
Feature
Propagation
01-1
De
9t-2
fmotion 9-1
fmotion Qt
Xt-2
xt-1
Our conditional coding-based framework
we
xt
Enigh
Feature
y
ヴィー
Extractor
AD
Rt
sec
Daigh
Diow
Fr
wder
Ct
Cr
Frame coding function
91-1
Entropy
2
Model
C
Xt
frontext
F-1-
Feature
Extractor
Motion
Alignment
Context
Rebesh
Control
P
Temporal feature modulation
5+
Results
CVPR
SEATTLE, WA JUNE 17-21, 2024
3000
High bitrate scenario
600
Target bitrate
Actual bitrate
Low bitrate scenario
2500
500-
1500
1000
100
200
300
Frame Index
300
Target bitrate
Actual bitrate
200
200
Frame index
Rate control examples
UVG MCL-JCV
VTM-17.0 0.0
0.0
HM-16.25 40.1
48.61
HEVC B
HEVC C HEVC D
HEVC E
Average
0.0
0.0
0.0
0.0
0.0
47.6
41.0
34.5
42.8
42.4
ECM-5.0 -14.9
-17.0
-17.3
-16.6
-16.1
-14.1
-16.0
ECM-11.0 -20.0
-22.1
-22.2
-21.2
-20.4
-17.2
-20.5
DCVC-DC 5.7
-5.0
12.2
-4.2
-16.5
84.4
12.8
DCVC-FM -20.7
-10.3
-18.2
-32.2
-41.2
-30.2
-25.5
BD-Rate (%) comparison, and the lower, the better
Original
ECM-5.0
DCVC-DC
DCVC-FM
BPP/PSNR
0.0184/32.74
0.0184/31.96
0.0178/33.36
BAY QUARTER
BAY QUARTER
BAY QUARTER
BAY QUARTER
BPP/PSNR
0.0183/35.22
0.0180/36.37
0.0180/34.93
Visual comparison