2022-08-08, Oral#2
MON-O-05, DiffusionCLIP: Text-Guided Diffusion Models for Robust Image Manipulation (예종철, KAIST)
CLIP (Radford, ICML 2021)
GAN Inversion
StyleCLIP -> StyleGAN-NADA (Siggraph 2021)
◆ Diffusion CLIP
– Not using DDPM
– DDIM ( Simplified version of DDPM )
하나의 Domain 마다 Fine-tuning 된 모델 있어야…
MON-O-06, GAN Inversion for out-of-range images with geometric transformations (조성현, POSTECH)
◆ Problems
– Dataset-Bias
– Out-of-range images
① BD Invert
– Base code : geometric transformation
– Detail code :
Base code 는 Convolution Layer 뒤에 적당한 feature map 을 빼내서 사용
② Regularized Optimization Scheme (좋은 content code)
-> encoder for base code
-> regularization for detail code
reconstruction 을 좋게 하는데 초점을 맞춤 <-> Editing 성능은 좋지 않음
Industry #1
1) Lunit
① 100,000 X 100,000 pixels ( huge data ) -> large images
② Fine-grained annotations required
③ Many sources of variability across images
annotation “CD3DET”
Barlow : Twin
2) 스트라드비전
“Software-defined car”
EDU -> DCU ( Domain Control using multi-channel cammera )
Invited Talk #1
Antonio Torralba (MIT)
“Learning to see by looking at noise”
-> Most important is sensing
GAN Dissection -> Style GAN
* Layer4 Neuron 119 ‘tree’
* Layer4 Neuron 43 ‘dome’
DataSetGAN
BigDatasetGAN
Pyramid based ~
Shades21K dataset
Images Labels
2022-08-09, Oral#3
TUE-O-01 Video-Question Answering Using Language-Guided Deep Compressed-Domain Video Feature
‘Video Compression Features’
Video Coding ?
– Intra-coding (I-Frame)
– P-B frame (motion vector & residual)
TUE-O-02 ‘Meta-Learning Sparse Implicit Neural Representations (INR)’, 신진우 KAIST
- INR, f(X,Y) = RGB
장점 : SR: Novel-View, Video Representation, Scalable
Obstables : Cost of Training
Alternative : Sparse Initial model (less parameter) + Good Initialization
Meta Learning + Network Prunning
Recent Works> Shifting modulation, l_0 regularization
2022-08-10, Oral#5
WED-O-02 ‘End-to-End Trainable Trident Person Search Network Using Adaptive Gradient Propagation’ (심재영)
- Person Search = Person(pedestrian) Detection + Person re-ID
기존: Two-step, End-to-End
문제점 :
1) Task Conflict
– P.D. -> person commonness
– P. re-ID -> uniqueness for each identity
2) Insufficient features
AGWF (Adaptive Gradient Weight Function)
: Detection 의 Confidence (αi) 를 weight 로 사용
-> Detection 이 잘 안 되면 consine similirity 활용도가 떨어진다는 가정
-> Detection 이 잘 된 Bounding Box 를 re-ID 하면 더 잘 된다.
Invited Talk #3
WED-I Chelsea Finn, “Robust Deep Networks through Invariance and Adaptation”
Oral#6
WED-O-05, XVFI : eXtreme Video Frame Interpolation (김문철, KAIST)
문제점 :
– Occlusion
– Deformation
– Large motion
WED-O-06, “Contrastive Learing for Space-Time Correspondence via Self-cycle Consistency” (Jeany Son)
Supervised? No!
Self-Supervised: forward —- backward (cycle-consistency)
supervision -> affinity matrix
self-supervision -> contrastive learning
문제점 : occlusion 문제로 propagation 중단 됨
해결책 :
1) self-cycle edges
2) bayesian model averaging (handling multi-hop)
Industry#3
- NaverLABs
– Computer Vision for in / outdoor mobility
Thanks.
I appreciate you.