Methods for improving speed and accuracy of single stage object detectors
The presentation is focused on object detectors training from random weights, i.e. without using pre-trained feature extractor. It considers the advantages and disadvantages of the approach.
The speech contains a lot of experiments with feature extractor and detector’s head to obtain tradeoff between speed and accuracy.
GroupNormalization is now a popular trend in Convolutional Neural Networks. It is shown when this approach gives a really great performance.
Methods which have large accuracy require an incredibly big amount of video memory. For this purpose, tricks to decrease memory for training process are described.
Filter amount in each convolutional layer is heuristic value. The presentation shows how one can find an optimal rank of the weight tensors for the problem.