In this paper, we investigate whether fusing depth information on top of normal RGB data for camera-based object detection can help to increase the performance of current state-of-the-art single-shot detection networks. Indeed, depth sensing is easily acquired using depth cameras such as a Kinect or stereo setups. We investigate the optimal manner to perform this sensor fusion with a special focus on lightweight single-pass convolutional neural network (CNN) architectures, enabling real-time processing on limited hardware. For this, we implement a network architecture allowing us to parameterize at which network layer both information sources are fused together. We performed exhaustive experiments to determine the optimal fusion point in the network, from which we can conclude that fusing towards the mid to late layers provides the best results. Our best fusion models significantly outperform the baseline RGB network in both accuracy and localization of the detections.

In this paper, we investigate the feasibility of automatic small object detection, such as vehicles and vessels, in satellite imagery with a spatial resolution between 0.3 and 0.5 m. The main challenges of this task are the small objects, as well as the spread in object sizes, with objects ranging from 5 to a few hundred pixels in length. We first annotated 1500 km², making sure to have equal amounts of land and water data. On top of this dataset we trained and evaluated four different single-shot object detection networks: YOLOV2, YOLOV3, D-YOLO and YOLT, adjusting the many hyperparameters to achieve maximal accuracy. We performed various experiments to better understand the performance and differences between the models. The best performing model, D-YOLO, reached an average precision of 60% for vehicles and 66% for vessels and can process an image of around 1 Gpx in 14 s. We conclude that these models, if properly tuned, can thus indeed be used to help speed up the workflows of satellite data analysts and to create even bigger datasets, making it possible to train even better models in the future.

In this paper we investigate the benefit of using depth information on top of normal RGB for camera-based pedestrian detection. Indeed, depth sensing is easily acquired using depth cameras such as a Kinect or stereo setups. We investigate the best way to perform this sensor fusion with a special focus on lightweight single-pass CNN architectures, enabling real-time processing on limited hardware. We implement different network architectures, each fusing depth at different layers of our network. Our experiments show that midway fusion performs the best, outperforming a regular RGB detector substantially in accuracy. Moreover, we prove that our fusion network is better at detecting individuals in a crowd, by demonstrating that it has both a better localization of pedestrians and is better at handling occluded persons. The resulting network is computationally efficient and achieves real-time performance on both desktop and embedded GPUs.

Object detection models are usually trained and evaluated on highly complicated, challenging academic datasets, which results in deep networks requiring lots of computations. However, a lot of operational use-cases consist of more constrained situations: they have a limited number of classes to be detected, less intra-class variance, less lighting and background variance, constrained or even fixed camera viewpoints, etc. In these cases, we hypothesize that smaller networks could be used without deteriorating the accuracy. However, there are multiple reasons why this does not happen in practice. Firstly, overparameterized networks tend to learn better, and secondly, transfer learning is usually used to reduce the necessary amount of training data. In this paper, we investigate how much we can reduce the computational complexity of a standard object detection network in such constrained object detection problems. As a case study, we focus on a well-known single-shot object detector, YoloV2, and combine three different techniques to reduce the computational complexity of the model without reducing its accuracy on our target dataset. To investigate the influence of the problem complexity, we compare two datasets: a prototypical academic (Pascal VOC) and a real-life operational (LWIR person detection) dataset. The three optimization steps we exploited are: swapping all the convolutions for depth-wise separable convolutions, perform pruning and use weight quantization. The results of our case study indeed substantiate our hypothesis that the more constrained a problem is, the more the network can be optimized. On the constrained operational dataset, combining these optimization techniques allowed us to reduce the computational complexity with a factor of 349, as compared to only a factor 9.8 on the academic dataset. When running a benchmark on an Nvidia Jetson AGX Xavier, our fastest model runs more than 15 times faster than the original YoloV2 model, whilst increasing the accuracy by 5% Average Precision (AP).