![]() Robinet, François ![]() ![]() ![]() in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2022 (2022, January 04) Free space estimation is an important problem for autonomous robot navigation. Traditional camera-based approaches train a segmentation model using an annotated dataset. The training data needs to capture ... [more ▼] Free space estimation is an important problem for autonomous robot navigation. Traditional camera-based approaches train a segmentation model using an annotated dataset. The training data needs to capture the wide variety of environments and weather conditions encountered at runtime, making the annotation cost prohibitively high. In this work, we propose a novel approach for obtaining free space estimates from images taken with a single road-facing camera. We rely on a technique that generates weak free space labels without any supervision, which are then used as ground truth to train a segmentation model for free space estimation. Our work differs from prior attempts by explicitly taking label noise into account through the use of Co-Teaching. Since Co-Teaching has traditionally been investigated in classification tasks, we adapt it for segmentation and examine how its parameters affect performances in our experiments. In addition, we propose Stochastic Co-Teaching, which is a novel method to select clean samples that leads to enhanced results. We achieve an IoU of 82.6%, a Precision of 90.9%, and a Recall of 90.3%. Our best model reaches 87% of the IoU, 93% of the Precision, and 93% of the Recall of the equivalent fully-supervised baseline while using no human annotations. To the best of our knowledge, this work is the first to use Co-Teaching to train a free space segmentation model under explicit label noise. Our implementation and trained models are freely available online. [less ▲] Detailed reference viewed: 116 (23 UL)![]() Pinel, Frédéric ![]() ![]() in Communications in Computer and Information Science (2020, February) We present a procedure for the design of a Deep Neural Net- work (DNN) that estimates the execution time for training a deep neural network per batch on GPU accelerators. The estimator is destined to be ... [more ▼] We present a procedure for the design of a Deep Neural Net- work (DNN) that estimates the execution time for training a deep neural network per batch on GPU accelerators. The estimator is destined to be embedded in the scheduler of a shared GPU infrastructure, capable of providing estimated training times for a wide range of network architectures, when the user submits a training job. To this end, a very short and simple representation for a given DNN is chosen. In order to compensate for the limited degree of description of the basic network representation, a novel co-evolutionary approach is taken to fit the estimator. The training set for the estimator, i.e. DNNs, is evolved by an evolutionary algorithm that optimizes the accuracy of the estimator. In the process, the genetic algorithm evolves DNNs, generates Python-Keras programs and projects them onto the simple representation. The genetic operators are dynamic, they change with the estimator’s accuracy in order to balance accuracy with generalization. Results show that despite the low degree of information in the representation and the simple initial design for the predictor, co-evolving the training set performs better than near random generated population of DNNs. [less ▲] Detailed reference viewed: 172 (15 UL)![]() Robinet, François ![]() ![]() in 2020 IEEE 17th Annual Consumer Communications & Networking Conference (CCNC) (2020) Convolutional Neural Networks have been successfully used to steer vehicles using only road-facing cameras. In this work, we investigate the use of Privileged Information for training an end-to-end lane ... [more ▼] Convolutional Neural Networks have been successfully used to steer vehicles using only road-facing cameras. In this work, we investigate the use of Privileged Information for training an end-to-end lane following model. Starting from the prior assumption that such a model should spend a sizeable fraction of its focus on lane markings, we take advantage of lane geometry information available at training time to improve its performance. To this end, we constrain the class of learnable functions by imposing a prior stemming from a lane segmentation task. For each input frame, we compute the set of pixels that most contribute to the prediction of the model using the VisualBackProp method. These pixel-relevance heatmaps are then compared with ground truth lane segmentation masks. A Distraction Loss term is added to the objective to regularize the training process. We learn from real-world data collected using our experimental vehicle and compare the results to those obtained using the simple Mean Squared Error objective. We show that the presence of our regularizer benefits both the performance and the stability of the model across a variety of evaluation metrics. We use a pretrained lane segmentation model without fine-tuning to extract lane marking masks and show that valuable learning signals can be extracted even from imperfect privileged knowledge. This method can be implemented easily and very efficiently on top of an existing architecture, without requiring the addition of any trainable parameter. [less ▲] Detailed reference viewed: 214 (46 UL) |
||