Announcing the Winners of the Joint COCO and Mapillary Recognition Challenge Workshop for ECCV 2018
Mapillary participates in several benchmarking challenges every year. This year, for instance, we’ve participated in (and won!) benchmarking challenges at both CVPR and ECCV. This is a great way for us to keep pushing the boundaries of what’s possible in the computer vision field and constantly evolve our own research and technology. We know how much added value these workshops have, so we’re very happy to have co-hosted a benchmarking challenge ourselves with COCO for this year’s ECCV.
As many of you know, COCO’s workshops have attracted hundreds of participants over the past couple of years. Because of its contributions to the Object Recognition community, the provided datasets & challenges, and participation from strong research labs, its workshops usually attract the best of the best.
This year, COCO and Mapillary co-held the Joint Recognition Challenge Workshop to advance the field of object recognition in the context of scene understanding. Mapillary and COCO both provided their datasets and were striving to unify task definitions and challenge guidelines, so each dataset and its characteristics could let researchers address different parts of object recognition tasks.
Object Detection Samples from the COCO dataset
COCO’s dataset is designed to further object detection research while focusing on a full scene understanding of general image scenes. Mapillary’s Vistas Dataset, on the other hand, has an emphasis on semantic image understanding of street scene environments, with applications for robot navigation or autonomous driving in mind. The two datasets complement each other in that COCO focuses on recognition in natural scenes, while Mapillary focuses on recognition of street-view scenes.
The workshop was split into four different tasks - two running on Mapillary’s dataset, and all four on COCO’s. COCO ran tasks in Object Detection, Keypoints Detection, Stuff Segmentation, and Panoptic Segmentation, with Mapillary running tasks on Object Detection and Panoptic Segmentation.
Object Detection Task Sample from the Mapillary Vistas Dataset
What’s different from previous years is the Panoptic Segmentation Challenge. That’s an entirely new challenge, designed to simultaneously address both stuff (material like grass, sky, or roads) and thing (countable objects, like humans, animals, vehicles) classes. These are normally kept separate, so this constitutes a significant departure from previous years’ workshops.
Panoptic Segmentation Sample from the Mapillary Vistas Dataset
Beyond the Panoptic Challenge, COCO’s DensePose Challenge is another new addition to this year’s workshop.
Here are the results:
The winners were announced at ECCV yesterday. Megvii, the facial recognition company from China, achieved the best score in four out of six tasks (COCO Detection, COCO Panoptic, COCO Keypoints, Mapillary Panoptic). This is, of course, very impressive and our warmest congratulations go out to the team. Other winners are Didi Map Vision (Mapillary Detection), MMDet (COCO Detection), and BUPT-PRIV (COCO DensePose). The runner-ups were MRSA (COCO Keypoints), Caribbean (COCO Panoptic), PKU_360 (COCO Panoptic), and TRI-ML (Mapillary Panoptic). Congratulations on all of your hard work!
The winners all did talks at ECCV, alongside PlumSix, ML_LAB, and Sound of Silent, all of whom presented particularly interesting research findings. We were also proud to host Professor Andreas Geiger with an invited talk, who was recently awarded the prestigious IEEE PAMI Young Researcher Award for his contributions to the computer vision field.
It’s been a joy co-organizing this workshop with COCO - huge congrats to the winners and everyone else participating in the workshop. Mapillary is at ECCV for the duration of the conference - you can find us at booth number 32.
/Peter Kontschieder, Head of Research at Mapillary
We would also like to thank our partners AWS and Nvidia. Thanks to them, both winners in the Mapillary tasks were awarded $10.000 in AWS credits and a Titan Xp GPU. Thank you for demonstrating your support for the research community that helps push the boundaries of what’s possible in the computer vision field.