Real Time Object Detection With A Single Network

Summary & Methodology Analysis

The paper addresses the need for fast and accurate algorithms for object detection by reframing detection as a regression problem to spatially separated bounding boxes and associated class probabilities. This enables a single neural network, known as YOLO, to predict detections directly from full images in one evaluation. The methodology divides the input image into an S x S grid where each grid cell predicts B bounding boxes, confidence scores, and C conditional class probabilities directly from image pixels. The network architecture is inspired by the GoogLeNet model for image classification, utilizing 1x1 reduction layers followed by 3x3 convolutional layers. The convolutional layers are pretrained on the ImageNet classification task at half resolution using the first 20 convolutional layers followed by an average-pooling layer and a fully connected layer. The Darknet framework is used for all training and inference. For detection, four convolutional layers and two fully connected layers with randomly initialized weights are added to the pretrained network, and input resolution is increased from 224 x 224 to 448 x 448. Optimization relies on a multi-part sum-squared error loss function that weights localization error, coordinate predictions, confidence predictions for cells containing or not containing objects, and conditional class probabilities using parameters such as lambda_coord and lambda_noobj. Combining Fast R-CNN with YOLO increases Fast R-CNN mAP on the VOC 2007 test set by 3.2% from 71.8% to 75.0%, and YOLO scores 57.9% mAP on the VOC 2012 test set. Despite these results, the approach has several limitations. YOLO imposes strong spatial constraints on bounding box predictions since each grid cell only predicts two boxes and can only have one class, which limits the number of nearby objects the model can predict. The model struggles with small objects that appear in groups, such as flocks of birds. Because the model learns to predict bounding boxes from data, it struggles to generalize to objects in new or unusual aspect ratios or configurations. It also uses relatively coarse features for predicting bounding boxes due to multiple downsampling layers from the input image. Finally, the loss function treats errors the same in small bounding boxes versus large bounding boxes, even though a small error in a small box has a much greater effect on IOU, making incorrect localizations the main source of error.