Hello! I am Valentin Khrulkov, a researcher from the Yandex Research team. We regularly attend industry conferences, and then share our impressions on Habré: which of the speakers was remembered, which stands could not be ignored, whose posters attracted the most attention. 2020 made significant adjustments to the usual schedule: many events were canceled and rescheduled, but the organizers of some of them risked trying new formats.
CVPR 2020 is 7600 participants, 5025 works, events and interactions, 1,497,800 minutes of discussions - and everything is online. More details are under the cut.
How it was: plans vs reality
Conference on Computer Vision and Pattern Recognition — — . : , — , «». CPRV — A, A1 A* — , ICCV ECCV. : 20% . — : , . — 25% . 6656 , — 1470, — 335. — , , . -.
: — , — , , — Zoom. , , , - -. , , . , :
- -: 350–500$. , ICLR, , 100$.
- . : - - .
- : , , , .
- : . , 12 : , . . , .
cvpr20.com:
: VS
: — , — . : , , ?
, , -: , . - : ? — . Zoom, .
Efficient Data Annotation for Self-Driving Cars via Crowdsourcing on a Large-Scale , 70 — - , - . ( , , , ). : . , , : , .
- , , — - , - , . , , .
, . , . , : 2D- () 3D- ( , ). , SD , . : , .
, , . — , 100 ( , ). — 30 000 , . : , «» — , . :
! , . Kitware, IBM Cornell, , .
Cross-Batch Memory for Embedding Learning:
: , . - GPU : .
: , , , , , / . ( ) .
CNN-generated images are surprisingly easy to spot… for now:
( GAN’) , : . , state-of-the-art GAN’, , .
: , . , , GAN, , .
Learning Better Lossless Compression Using Lossy Compression: ,
(lossless). : state-of-the-art (lossy), lossless .
, lossless . , CVPR , , , lossless-JPEG2000 WebP.
Image Processing Using Multi-Code GAN Prior:
, GAN , , ? , ?
, : , , Prior GAN — , , .
Effectively Unbiased FID and Inception Score and where to find them: GANs
FID . , .
, . , , , , .
FDA: Fourier Domain Adaptation for Semantic Segmentation:
domain adaptation: - - . , /. .
Single-Image HDR Reconstruction by Learning to Reverse the Camera Pipeline:
HDR (High Dynamic Range) LDR (Low Dynamic Range) : , , . LDR-HDR . : state-of-the-art, , .
A Multigrid Method for Efficiently Training Video Models: tradeoff
, , : , . , .
, : , , , . ( «», ). : 4,5 .
Towards Robust Image Classification Using Sequential Attention Models:
DeepMind ( CNN, LSTM ) — ( ) () . :
- Feedforward : , .
- «» : . , .
LSTM, ResNet': , — , — .
150 TPU , , . , «» , - .
Adversarial Vertex Mixup: Toward Better Adversarially Robust Generalization:
. generalization gap adversarial training ( 40%). :
- Mixup — .
- Label smoothing — - one-hot target, u: — u, (1 — u) .
. x x* ( PGD ), «adversarial vertex kx*», , x. , mixup (x kx*).
— ( ) (, , ).
, — SOTA — , PGD .
High-Resolution Daytime Translation Without Domain Labels:
, , ? , . , unsupervised .
: AdaIN , (style transfer) skip-connections, . , (1024 x 1024) : .
NLP 2017 , . , — , — : «-»? : — , — , .
Yandex Research , : ? : . , , : 10%. — .
, CVPR 2020 - . , , .