Git Product home page Git Product logo

kaggle-cvpr-2018-wad-video-segmentation-challenge-solution's Introduction

CVPR 2018 WAD Video Segmentation Challenge Solution

Overview

The goal of this Kaggle Challenge is to accurately segment objects such as car and pedestrian at instance-level from the video sequences recorded by the autonomous driving car. Our solution is based on Mask R-CNN. We use both Apolloscape and Cityscape datasets and only focus on a particular field of view in the image at the highest possible resolution. Our solution reaches the 2nd palce in the challenge (Private Score 0.30219).

Statistical Analysis

The image resolution of Apolloscape is 3384 by 2710. However, our interesting objects are not uniformly distributed over the image. Actually, as revealed in the 2D histogram below, the distribution is densely concentrated in a narrow stripe of the whole field of view. For example, between y = 1560 to y = 2280, 99.7% of our interesting objects are localized in this region of interest (ROI). This is our key observation.

Left, our perdiction result example. Right, 2D histogram of annotations.

Implemenation

The core algorithm in our solution is Mask R-CNN. There are a few open source implementations on the shelf. We use Detectron, the object detection library developed by the Mask R-CNN original authors Kaiming He, Piotr Dollar, Ross Girshick from Facebook AI Research team. As we learned from the previous Kaggle image segmentation competition, to achieve better performance, higher image resolution for training and inference is always preferred. However, due to the GPU memory size constraint and our statistical analysis in the previous section, we crop and resize the image to 3384 by 720. With a Nvidia P100 graphical card, training a 9-class Mask R-CNN FPN model with ResNet X101 backbone network requires around 1560 MB GPU memory when the batch size is 1. However, to have a reliable solution, a full field of view prediction at a lower resolution should be considered to recognize those objects that are not in our ROI coverage. In practice, we find it does not improve the performance. Another interesting finding is the threshold for the object detection. Through experiment, 0.25 is the optimal threshold to obtain the highest leader board score. We use the pretrained model from COCO dataset provided by the Detectron Model Zoo. More detailed information can be found from our project report. We also provide our model used for the competition.

kaggle-cvpr-2018-wad-video-segmentation-challenge-solution's People

Contributors

computational-camera avatar

Stargazers

 avatar

Watchers

 avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.