Using Deep Teaching, Helm.ai placed #1 on all metrics of the CVPR Lane Detection Challenge, an open autonomous driving benchmark held during the 2018 Conference on Computer Vision and Pattern Recognition.
We trained a neural network without human annotation or simulation on 10M+ images of dashcam footage from across the world. We then fine-tuned this network on the CVPR dataset and placed #1 on the leaderboard on June 16th 2018, using the anonymized username dpantoja for all our submissions.
Unlike the typically highly specialized efforts used in deep learning competitions, Helm.ai spent minimal engineering time to top out the leaderboard. As of August 2018 we ranked #1 in each category of scoring: accuracy, false positives and false negatives. Using mostly fine-tuning methods we maintained our lead through the closing of the competition in mid-September 2018.
Deep learning competitions capture only a tiny fraction of the validation of robustness required for autonomous driving, but the efficacy of the relatively “effortless” approach of fine-tuning from large scale training shows the accuracy and development speed advantages of Deep Teaching.
Stay tuned for an upcoming whitepaper from Helm.ai with more info!
