Youku Video Enhancement and Super‑Resolution Competition Overview

Video enhancement and super‑resolution are core computer‑vision algorithms that aim to restore degraded video content and improve visual clarity. The competition provides a platform for showcasing state‑of‑the‑art methods and offers the industry‑largest, most diverse dataset, covering various content categories, noise models, and difficulty levels.

Prize Structure: Champion (¥100,000), Runner‑up (¥60,000), Third place (¥40,000), Geek Award (¥10,000 ×3). The top‑10 teams in the semifinals gain access to Alibaba (Youku) campus recruitment green channel.

Schedule: Preliminary (May 5 – June 18, UTC+8), Semifinal (June 25 – August 10, UTC+8), Final (September, UTC+8).

Registration: Opens April 11, 2019 (UTC+8); deadline June 16, 2019 12:00 (UTC+8). Teams of 1‑5 members; each participant may join only one team. Registration is done via a Taobao or Alibaba Cloud account on the Tianchi website.

Competition Task: Build models that map low‑resolution video frames to high‑resolution frames (scale factor 1:4). Low‑resolution videos are generated by simulating real‑world noise and compression artifacts; high‑resolution videos come from Youku’s HD media library.

The dataset contains over 10,000 video pairs, with the first 1,000 released for the competition, ~2,000 to be publicly released after the contest, and the remaining 7,000 gradually opened.

Task Analysis: Besides upscaling, participants must address noise, compression artifacts, and other degradations. References include the “A collection of high‑impact and state‑of‑the‑art SR methods” repository, NTIRE competition reports, and baseline networks such as EDSR‑PyTorch ( https://github.com/thstkdgus35/EDSR-PyTorch ).

EDSR, the NTIRE 2017 champion, removes unnecessary residual modules for low‑level super‑resolution tasks and introduces a multi‑scale model with shared parameters. Recent advances like Meta‑SR enable arbitrary scaling factors via a Meta‑Upscale Module that predicts filter weights dynamically.

Additional relevant literature: Google’s RAISR ( https://arxiv.org/abs/1606.01299 ), NTIRE 2018 SR report, and papers on video denoising and compression‑artifact removal (e.g., Multi‑level Wavelet‑CNN, Residual Dense Network, Convolutional Auto‑encoders with Symmetric Skip Connections).

Video enhancement techniques from past NTIRE challenges, such as Meitu’s Range Scale Global‑UNet and Modified HRNet, are also recommended.

The competition aims to bridge academic research and industrial application, fostering collaboration and advancing video enhancement and super‑resolution technologies.