MergeNet: A Deep Net Architecture for Small Obstacle Discovery

Abstract

We present here, a novel network architecture called MergeNet for discovering small obstacles for on-road scenes in the context of autonomous driving. The basis of the architecture rests on the central consideration of training with less amount of data since the physical setup and the annotation process for small obstacles is hard to scale. For making effective use of the limited data, we propose a multi-stage training procedure involving weight-sharing, separate learning of low and high level features from the RGBD input and a refining stage which learns to fuse the obtained complementary features. The model is trained and evaluated on the Lost and Found dataset and is able to achieve state-of-art results with just 135 images in comparison to the 1000 images used by the previous benchmark. Additionally, we also compare our results with recent methods trained on 6000 images and show that our method achieves comparable performance with only 1000 training samples.

A Novel Multi-layer Framework for Tiny Obstacle Discovery

“…Similarly, when FPR is lower, our method achieves considerable improvement in accuracy over these two methods. MergeNet [3] utilizes deep learning to discover obstacles, and achieves an accuracy of 85% when FPR is 2.0%. Although Our method is not based on deep learning, it achieves an approximate result.…”

“…In general, there are three main categories for visual obstacle discovery: the correlation-based methods [1][2], the segmentation-based methods [3] [4], and the proposalbased methods [5] [6]. The first type compares the relative positions between 3D points in disparity map, and classifies all points into obstacle and road.…”

“…The second type segments an RGB image into several regions with different semantics. MergeNet [3] proposes a network architecture for discovering obstacles, which makes effective use of the limited data. However, tiny obstacles provide little information throughout the image, hence it is hard to discover them.…”

Tiny Obstacle Discovery by Occlusion-Aware Multilayer Regression

“…Table VIII depicts the comparison of time per image between our method and the previous method. Since [1], [2] does not provide complete information about the computing platform, it is difficult to make a comprehensive comparison. Our method is implemented in MATLAB and tested on a PC with 16 GB memory and an Intel Core i7-2600K CPU, and takes more time to process an image.…”

“…To alleviate the second issue mentioned above, a multistride sliding window scheme is proposed, i.e., a dense sampling proposal in the highlayer region, and a sparse sampling proposal in the low-layer region. Furthermore, to achieve a high discovery performance, we explore two appearance properties of obstacles: (1) the dissimilarity between the obstacle and road plane, (2) the dissimilarity between the obstacles and other objects. Then, an obstacle-aware regression model is designed to learn these dissimilarities and generate an obstacle-occupied probability map, which consists of a pair of primary and secondary regressors.…”