We are returning to the International Conference on Computer Vision (ICCV) this year to meet with fellow researchers and unveil some of our latest research. We’ve also taken part in four benchmarking challenges for this year’s ICCV. Each challenge is part of the Autonomous Navigation in Unconstrained Environments Workshop (AutoNUE), and we are very happy to say that we won all four challenges in the workshop, named Semantic Segmentation, Panoptic Segmentation, Segmentation on Constrained Devices, and Localization.

The competition was fierce and our results come down to how we optimize our models for solving each of the different tasks. We gained a headstart over other submissions by pre-training our image recognition models on the Mapillary Vistas Dataset. This led our models to generalize better on the test data, which is always what we’re after since we want our algorithms to perform well on fresh imagery. The Vistas Dataset is one of the most diverse datasets available for training algorithms to see and understand street-scenes. The dataset consists of imagery from more than 190 countries that has been captured and uploaded to Mapillary by people and companies everywhere. Training our algorithms on the Vistas Dataset has served us well, and has helped us secure numerous wins at benchmarking challenges at both CVPR and ICCV this year alone.

Come and find us at ICCV

We will attend the award ceremony for AutoNUE on November 2, where we will also give a talk about our winning submissions.

Images from the Mapillary Vistas Dataset

We are also very happy to once again host a Joint Recognition Challenge Workshop with the COCO Consortium that consists of members from organizations like Facebook AI Research, Google Brain, and others. We have invited research teams to benchmark their computer vision algorithms on the COCO and Mapillary Vistas datasets as part of our efforts to push the boundaries of the field. Several leading research teams are taking part, and we will unveil the results at ICCV.

We have lots more activities lined up. See below to get an overview of where you’ll find us.

We will also be at booth E31 throughout the week, and we would love to see you! Come by and play the Mapillary Verifications Projects to help improve our computer vision algorithms, see what your hometown looks like through the eyes of an autonomous vehicle, and, of course, speak to our Computer Vision team about our latest wins and what we have coming up.

/Peter, Director of Research