ATOM: Accurate Tracking by Overlap Maximization This paper proposes a novel tracking architecture for visual object tracking, consisting of two components: a target estimation module trained offline and a target classification module trained online. The target estimation module predicts the Intersection over Union (IoU) overlap between the target and an estimated bounding box, while the classification module is trained online to provide high discriminative power in the presence of distractors. The proposed method achieves state-of-the-art performance on five challenging benchmarks, including TrackingNet, where it achieves a relative gain of 15% over the previous best approach. The tracker runs at over 30 FPS and is implemented in Python using PyTorch. The target estimation component is trained offline on large-scale datasets to predict the IoU overlap between the target and a bounding box estimate. It integrates target-specific knowledge by performing feature modulation. The classification component consists of a two-layer fully convolutional network head and is trained online using a dedicated optimization approach. The method is evaluated on five challenging tracking datasets: Need for Speed (NFS), UAV123, TrackingNet, LaSOT, and VOT2018. On the TrackingNet dataset, the tracker achieves a relative gain of 15% in terms of success. On the LaSOT dataset, it achieves an absolute gain of 10% in success. On the VOT2018 dataset, it achieves the best EAO score of 0.401, with a relative gain of 3% over LADCF. The method outperforms previous methods in terms of precision, normalized precision, and success on the TrackingNet dataset. It also outperforms previous methods in terms of expected average overlap (EAO), robustness (tracking failure), and accuracy on the VOT2018 dataset. The method is robust against distractor objects in the scene and provides accurate target estimation. The proposed approach is effective in handling challenging tracking scenarios, including deformation, view change, occlusion, fast motion, and distractors. The method is implemented with a focus on efficiency and performance, achieving high frame rates and accurate tracking results.ATOM: Accurate Tracking by Overlap Maximization This paper proposes a novel tracking architecture for visual object tracking, consisting of two components: a target estimation module trained offline and a target classification module trained online. The target estimation module predicts the Intersection over Union (IoU) overlap between the target and an estimated bounding box, while the classification module is trained online to provide high discriminative power in the presence of distractors. The proposed method achieves state-of-the-art performance on five challenging benchmarks, including TrackingNet, where it achieves a relative gain of 15% over the previous best approach. The tracker runs at over 30 FPS and is implemented in Python using PyTorch. The target estimation component is trained offline on large-scale datasets to predict the IoU overlap between the target and a bounding box estimate. It integrates target-specific knowledge by performing feature modulation. The classification component consists of a two-layer fully convolutional network head and is trained online using a dedicated optimization approach. The method is evaluated on five challenging tracking datasets: Need for Speed (NFS), UAV123, TrackingNet, LaSOT, and VOT2018. On the TrackingNet dataset, the tracker achieves a relative gain of 15% in terms of success. On the LaSOT dataset, it achieves an absolute gain of 10% in success. On the VOT2018 dataset, it achieves the best EAO score of 0.401, with a relative gain of 3% over LADCF. The method outperforms previous methods in terms of precision, normalized precision, and success on the TrackingNet dataset. It also outperforms previous methods in terms of expected average overlap (EAO), robustness (tracking failure), and accuracy on the VOT2018 dataset. The method is robust against distractor objects in the scene and provides accurate target estimation. The proposed approach is effective in handling challenging tracking scenarios, including deformation, view change, occlusion, fast motion, and distractors. The method is implemented with a focus on efficiency and performance, achieving high frame rates and accurate tracking results.