Single-object visual tracking aims at locating a target in each video frame by predicting the bounding box of the object. Recent approaches have adopted iterative procedures to gradually refine the bounding box and locate the target in the image. In such approaches, the deep model takes as input the image patch corresponding to the currently estimated target bounding box, and provides as output the probability associated with each of the possible bounding box refinements, generally defined as a discrete set of linear transformations of the bounding box center and size. At each iteration, only one transformation is applied, and supervised training of the model may introduce an inherent ambiguity by giving importance priority to some transformations over the others. This paper proposes a novel formulation of the problem of selecting the bounding box refinement. It introduces the concept of non-conflicting transformations and allows applying multiple refinements to the target bounding box at each iteration without introducing ambiguities during learning of the model parameters. Empirical results demonstrate that the proposed approach improves the iterative single refinement in terms of accuracy and precision of the tracking results.
28 March 2024 in Neural Computing and Applications
This paper proposes a new method to improve the training efficiency of deep convolutional neural networks. During training, the method evaluates scores to measure how much each layer’s parameters change and whether the layer will continue learning or not. Based on these scores, the network is scaled down such that the number of parameters to be learned is reduced, yielding a speed-up in training. Unlike state-of-the-art methods that try to compress the network to be used in the inference phase or to limit the number of operations performed in the back-propagation phase, the proposed method is novel in that it focuses on reducing the number of operations performed by the network in the forward propagation during training. The proposed training strategy has been validated on two widely used architecture families: VGG and ResNet. Experiments on MNIST, CIFAR-10 and Imagenette show that, with the proposed method, the training time of the models is more than halved without significantly impacting accuracy. The FLOPs reduction in the forward propagation during training ranges from 17.83% for VGG-11 to 83.74% for ResNet-152. As for the accuracy, the impact depends on the depth of the model and the decrease is between 0.26% and 2.38% for VGGs and between 0.4 and 3.2% for ResNets. These results demonstrate the effectiveness of the proposed technique in speeding up learning of CNNs. The technique will be especially useful in applications where fine-tuning or online training of convolutional models is required, for instance because data arrive sequentially.