2.3. Object Detection Methods

The object detection methods compared in this study are briefly described in the following (the following source codes were used as a basis for our implementation: Faster-RCNN, https://github.com/yhenon/keras-frcnn; YOLOv3, https://github.com/qqwweee/keras-yolo3; and RetinaNet, https://github.com/fizyr/keras-retinanet).

Faster-RCNN [30]: In this method, a feature map is initially produced by a ResNet50 [31]. Given the feature map, Faster-RCNN detects object instances in two stages. The first stage, called Region Proposal Network (RPN), receives the feature map and proposes candidate object bounding boxes. The second stage also accesses the feature map and extracts features from each candidate bounding box using a Region of Interest Pooling (RoIPoolRoIPool) layer. This operation is based on max pooling, and aims to obtain a fixed-size feature map, independent on the size of the candidate bounding box at its input. A softmax layer then predicts the class of the proposed regions as well as the offset values for their bounding boxes.

YOLOv3 [32]: Unlike Faster-RCNN, which has a stage for region proposal, YOLOv3 addresses the object detection as a problem of direct regression from pixels to bounding box coordinates and class probabilities. The input image is divided into S×S tiles. For each tile, YOLOv3 predicts bounding boxes using dimension clusters as anchor boxes [33]. For each bounding box, an objectness score is predicted using logistic regression, which indicates the chance of the bounding box to have an object of interest. In addition, C class probabilities are estimated for each bounding box, indicating the classes that it may contain. In our case, each bounding box may contain the cumbaru species or background (uninteresting object). Thus, each prediction in YOLOv3 is composed of four parameters for the bounding box (coordinates), one objectness score and C class probabilities. To improve detection precision, YOLOv3 predicts boxes at three different scales using a similar idea to feature pyramid networks [34]. As a backbone, YOLOv3 uses Darknet-53 as it provides high accuracy and requires fewer operations compared to other architectures.

RetinaNet [35]: Similar to YOLOv3, RetinaNet is a one-stage object detector but it addresses class imbalance by reducing the loss assigned to well-classified images. Class imbalance occurs when the number of background examples is much larger than examples of the object of interest (cumbaru trees). Using this new loss function, training focuses on hard examples and prevents the large number of background examples from hampering method learning. RetinaNet architecture consists of a backbone and two task-specific subnetworks (see Figure 1b). As the backbone, RetinaNet adopts the Feature Pyramid Network from [34], which is responsible for computing a feature map over an entire input image. The first subnet is responsible for predicting the probability of object’s presence at each spatial position. This subnet is a small Fully Convolutional Network (five conv layers) attached to the backbone. The second subnet, which is parallel with the object classification subnet, performs bounding box regression. The design of this subnet is identical to the first one except that it estimates box coordinates for each spatial location at the end.