Papers with Code Newsletter #1

🚀 Hello World!

Hi everyone,

Welcome to our new Papers With Code Newsletter. The objective of the newsletter is to provide you an overview of informative research highlights and trending papers with code around machine learning (ML) research, including progress, new tasks, benchmarks and datasets releases relevant to different areas of ML research.

Trending Papers with Code

Here are the trending papers with code for last week:

1. Papers with Official Code

Deformable DETR: Deformable Transformers for End-to-End Object Detection (paper & code) - Deformable DETR improves the performance of a previous approach, DETR, by introducing a new type of attention module that requires less training.
#1 trending with 434 ★

Navigating the GAN Parameter Space for Semantic Image Editing (paper & code ) - This work uses GANs to produce impressive visual effects thanks to the ability to control generation through discovery of interpretable directions of the latent space. This allows for non-trivial semantic manipulations and effective navigation of the GAN parameter space for semantic image editing.
#3 trending with 192 ★

Visual effects achieved by navigating the parameter space of StyleGAN2 on the LSUN-Horse dataset.

pixelNeRF: Neural Radiance Fields from One or Few Images (paper & code) - Recent method neural radiance field (NeRF) for novel view synthesis requires large number of images and lengthy per-scene optimization making it impractical. pixelNeRF is a learning framework for predicting neural radiance fields from one or few input images. The authors claim to outperform previous methods on novel view synthesis and single image 3D reconstruction.
#4 trending with 230 ★

Fully Convolutional Networks for Panoptic Segmentation (paper & code) - This work proposes Panoptic FCN which is a method for panoptic segmentation based on a fully convolutional pipeline. Authors claim a highly-efficient model applied to COCO, Cityscapes, and other vision datasets.
#6 trending with 96 ★

PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models (paper & code) - Single-image super resolution aims to construct high resolution images from low resolution inputs. PULSE is a self-supervised algorithm for super resolution that generates images that are both realistic and downscale correctly. This was work was published in CVPR 2020.
#7 trending with 4719 ★

Aligning Superhuman AI with Human Behavior: Chess as a Model System (paper & code) - This work introduces a variant of Alpha-Zero that predicts human moves more accurately than existing chess engines. The system is designed to accurately model granular human decision-making and propose to encourage collaboration between humans and algorithms.
#8 trending with 138 ★

TinaFace: Strong but Simple Baseline for Face Detection (paper & code) - TineFace is a simple baseline method for face detection based on existing generic object detection modules. While simplistic, it outperforms existing methods with larger backbone on a challenging face detection benchmark.
#9 trending with 192 ★

The model architecture of TinaFace

Weakly Supervised Person Re-ID: Differentiable Graphical Learning and A New Benchmark (paper & code) - Proposes a differentiable graphical model, capturing image dependencies, to generate pseudo labels that leads to a weakly supervised approach for re-identification. The authors use a trick to reduce annotation efforts needed and propose a new benchmark dataset for the task of re-identification called SYSU-30k.
#10 trending with 104 ★

2. Community Implementations

ArtLine (code & paper) - Self attention GAN (SAGAN) enables attention-driven, long-range dependency for image generation. ArtLine is a new library to easily create line art portraits using SAGANs, powered by the improved generator.
#2 trending 494 ★

DAIN (code & paper) - Video frame interpolation aims to synthesize non-existent frames in-between the original frames. DAIN is one method used to efficiently address this problem. DAIN-APP is a video-interpolation application which comes with a user interface that allows the DAIN technique to be used easily with minimal change to code.
#5 trending with 187 ★

Community Contribution Highlights

In this section we highlight some of the new community contributors to benchmarks, tasks and methods:

Drug-Drug Interaction Extraction - This task now has a new benchmark for drug-drug interaction extraction, with a method proposed in a recent paper by Asada et al getting the top results.
Thanks @Masaki_Asada for adding the results.

New SOTA on ImageNet - A new paper by Foret et al improves the state-of-the-art performance on ImageNet benchmark and obtain good results on a variety of other benchmarks. Authors propose an optimization procedure involving simultaneous minimization of loss value and loss sharpness with the goal to improve model quality and generalization.
Thanks @pforet for adding the results.

You Only Need Adversarial Supervision - A paper under review in ICLR2021 by Su et el proposes a simplified GAN model that relies only on adversarial supervision to improve quality of results on semantic image synthesis. Achieves top results on the Image-to-image Translation on ADE20K-Outdoor Labels-to-Photos benchmark.
Thanks @apata for adding the results.

Multimodal Unsupervised Image-To-Image Translation - is a task to produce multiple translations from one single image coming from a different domain. In relevant work, this recent paper by Or-El et al published at ECCV 2020, propose a novel multi-domain image-to-image generative adversarial network architecture as an aging transformation synthesis method.
Thanks @royorel for adding the task.

Other recent benchmark results added this past week by contributors:

@cpt - Added results for Doubly Stochastic Subspace Clustering
@kinyiu - Added results for Scaled-YOLOv4: Scaling Cross Stage Partial Network
@AR - Added several benchmark results for sentiment analysis, including for Gradient Harmonized and Cascaded Labeling for Aspect-based Sentiment Analysis
@sun008sun - Added results for Diverse Temporal Aggregation and Depthwise Spatiotemporal Factorization for Efficient Video Classification
@phoenix000 - Added results for FGN: Fusion Glyph Network for Chinese Named Entity Recognition
@cyou - Added results for Learning Diverse and Discriminative Representations via the Principle of Maximal Coding Rate Reduction

A huge thanks to @zhaosssss, @leoribeiro, @liminyan, @alexeyab84, @haiwen_huang, @HTVR and hundreds of other contributors for their multiple contributions to tasks, methods, benchmark results and code links.

Get in touch with elvis@paperswithcode.com to learn how to get your contributions highlighted. Optionally, you can connect with us on Slack (#newsletter channel).

Announcements and Final Words

Papers with Code is expanding to more sciences: physics, maths, CS, statistics and astronomy. Check out the new portal here. Read more about our efforts in this blog post and catch all our announcements on Twitter.

NeurIPS 2020 papers with code - NeurIPS 2020 happened last week! Thanks to all the authors who provided code for their work, we indexed 700 papers with code implementations. Note that on the Papers with Code website you can filter by task/author and sort by GitHub stars or alphabetically.

Finally, we would love to hear your feedback and suggestions for the newsletter. Just hit reply! If you want to know about how to get your papers and contributions featured, please contact us at elvis@paperswithcode.com.

Thanks for reading,
The Papers With Code Team