Minimizing Supervision for Free-Space Segmentation

Abstract

Identifying "free-space," or safely driveable regions in the scene ahead, is a fundamental task for autonomous navigation. While this task can be addressed using semantic segmentation, the manual labor involved in creating pixelwise annotations to train the segmentation model is very costly. Although weakly supervised segmentation addresses this issue, most methods are not designed for free-space. In this paper, we observe that homogeneous texture and location are two key characteristics of free-space, and develop a novel, practical framework for free-space segmentation with minimal human supervision. Our experiments show that our framework performs better than other weakly supervised methods while using less supervision. Our work demonstrates the potential for performing free-space segmentation without tedious and costly manual annotation, which will be important for adapting autonomous driving systems to different types of vehicles and environments.

Weakly-Supervised Free Space Estimation through Stochastic Co-Teaching

“…Bottom-Half is able to reach a decent IoU of 0.7550 and a high Recall, which is not surprising since free space indeed covers a large portion of the lower half of most frames. The Precision of this model is however only of 0.7798, which is poor compared to the 0.8778 achieved by our second unsupervised baseline, the raw Weak Labels from [49]. This second baseline also yields a large IoU improvement, reaching 0.7900.…”

“…We present Co-Teaching and its adaptation for a segmentation task, and we introduce its Stochastic variant. Since we focus on improving the training aspect, we use the weak labels proposed in [49] as targets during training. We benchmark the performances of (Stochastic) Co-Teaching against a fully-supervised model, as well as against unsupervised and weakly-supervised baselines in Section 5.…”

“…Network architectures Following recent research that shows that a well-tuned vanilla U-Net can outperform many variants on segmentation tasks [21], we opt for a U-Net structure based on a ResNet18 backbone (14.3M parameters) [47,18,55]. To compare with prior art, we also implement and train the SegNet model described in [49]. For computational reasons, we use a 512 × 1024 input resolution in all experiments.…”

“…Training procedure Since the authors of [49] do not share trained weights for their SegNet architecture, all models are trained from randomly initialized weights to allow for fair comparison. We use the PyTorch framework [43] and train models with the Adam optimizer [28], a batch size of 6 and a learning rate of 0.001.…”

“…This creates a need for very large datasets, and renders fully-supervised semantic segmentation solutions impractical. We tackle this problem in a different way: relying on a method that generates weak, noisy, free space annotations without any supervision [49], we train a neural network to generalize past the label noise using Co-Teaching [16].…”