A presenter at a conference explains a detailed research poster on the topic of "Task-Aware Encoder Control for Deep Video Compression." The individual is actively gesturing towards the poster, engaging the audience with key points of the research. The poster is filled with complex diagrams, charts, and text explaining methods such as "Controlling DVC for Machine" and "Dynamic Vision Mode Prediction (DVMP)." The presenter wears a lanyard with identification badges and stands in a dimly lit exhibition hall with overhead fluorescent lighting, featuring rows of similar posters. The atmosphere suggests a collaborative and educational event focused on advancements in technology and machine learning.
Text transcribed from the image:
Task-Aware Encoder Control for Deep Video Compression
Xingwei Gao, Zhihao Wang, Juncheng Li, Shuo Wang, Yue Wang, Zhibo Chen
University of Science and Technology of China

“Controlling DVC for Machine” Framework

Dynamic Vision Mode Prediction(DVMP)

Previous works require individually customized codecs to support different tasks, which is complex and difficult to deploy. 

However, our proposed unified video compression method can adaptively support various downstream tasks in a lightweight way through three key techniques:
1. Dividing the original P frame into two types: X frame and the other P-frames with DVMP.
2. Using GoP Selection Module to control the encoding GoP structure for different objectives, such as vision tasks and video reconstruction.
3. Maintaining the decoder weights constant to ensure compatibility across versions.

HyperPrll\  information

Reference  
input frames

Encod|ed -- Residual 

+oherence

Hyperprver RFT 

Effi

Efficienty  

ciency

Encod: Nestedioce

Meigcued  

CP,)

Encod: Nestedioce

Meigcued  

goP Coming

Bitstream

(Top) DVMP uses both PWC 
netflow tocope with

and DCVC toald 

(Down) DVMP ef acll
thenot gn and 
cpe wh semntie
w motion 

In hybrid cases of Fig4 structure

Loss Function & Training

Lf=L2+λldrate

Using perception-based and

Lf (similar for )

λr: rate-dist giesionaltsnniageenmder axis

Ov loss

λt: printisee research terms,

λD>0 ,λD <=1

λrec

(trancdentative)

(a)MPEG

(b)

(d)Insert_CMfor 

Rate-D -stortion

PSNR(dB)

For both subjective

T Dataset(lm) DVC-M-SS QP,

PSNR

&disori

SSIM(QP>40)

Task Rec

14.732

9.372

67.21%

PSNR(dB)

For both subjective

T Dataset(lm) DVC-M-SS QP,