Section 5.3
Object Detection Evaluation and Progress

Of course reporting speed is not enough, we also need to report the accuracy (i.e., mAP) of a detection model and there is often a trade-off between the two. Figure 63 shows some performance graphs reported in the literature plotting the trade-off between speed and accuracy for various model architectures (left) and increasing accuracy as models have been improved over the past several years.

Note that the performance of a detector depends on many factors, including the method, the backbone convolutional neural network architecture, image resolution, training regime, datasets used for training and evaluation, thresholds used for filtering, non-maximal suppression and IoU metrics, etc. Determining the best settings and combinations is a subject of ongoing research. Always be conscious of these factors when comparing models.