Spatial Transformer Networks (STNs) are a learnable module that enables neural networks to actively transform feature maps based on the input data. Introduced in this paper, STNs allow for spatial manipulation within the network, enabling models to learn invariance to transformations such as translation, scale, rotation, and more complex warping. This capability is achieved through a differentiable module that can be integrated into existing convolutional architectures without requiring additional training supervision or changes to the optimization process. The module consists of three components: a localisation network that predicts transformation parameters, a grid generator that creates a sampling grid, and a sampler that applies the transformation to the feature map. The localisation network takes the input feature map and outputs the parameters of the spatial transformation to be applied. The grid generator creates a sampling grid based on these parameters, and the sampler applies the transformation to the feature map. This process allows the network to focus on relevant regions of the input and transform them to a canonical pose, improving recognition in subsequent layers. STNs can be trained using standard back-propagation, enabling end-to-end training of the models they are injected into. Experiments show that STNs significantly improve performance on various tasks, including image classification, co-localisation, and fine-grained classification. For example, on distorted MNIST datasets, STNs outperform traditional CNNs by achieving lower error rates. On the Street View House Numbers (SVHN) dataset, STNs achieve state-of-the-art results in multi-digit recognition. In fine-grained classification on the CUB-200-2011 birds dataset, STNs achieve higher accuracy by learning to attend to specific object parts. STNs are also effective in semi-supervised scenarios, such as co-localisation, where they can localise common objects in a set of images without using object class labels or ground truth locations. The use of STNs allows for more efficient and accurate processing of images by focusing on relevant regions and transforming them to a canonical pose. Overall, STNs provide a flexible and powerful tool for improving the performance of neural networks in various computer vision tasks.Spatial Transformer Networks (STNs) are a learnable module that enables neural networks to actively transform feature maps based on the input data. Introduced in this paper, STNs allow for spatial manipulation within the network, enabling models to learn invariance to transformations such as translation, scale, rotation, and more complex warping. This capability is achieved through a differentiable module that can be integrated into existing convolutional architectures without requiring additional training supervision or changes to the optimization process. The module consists of three components: a localisation network that predicts transformation parameters, a grid generator that creates a sampling grid, and a sampler that applies the transformation to the feature map. The localisation network takes the input feature map and outputs the parameters of the spatial transformation to be applied. The grid generator creates a sampling grid based on these parameters, and the sampler applies the transformation to the feature map. This process allows the network to focus on relevant regions of the input and transform them to a canonical pose, improving recognition in subsequent layers. STNs can be trained using standard back-propagation, enabling end-to-end training of the models they are injected into. Experiments show that STNs significantly improve performance on various tasks, including image classification, co-localisation, and fine-grained classification. For example, on distorted MNIST datasets, STNs outperform traditional CNNs by achieving lower error rates. On the Street View House Numbers (SVHN) dataset, STNs achieve state-of-the-art results in multi-digit recognition. In fine-grained classification on the CUB-200-2011 birds dataset, STNs achieve higher accuracy by learning to attend to specific object parts. STNs are also effective in semi-supervised scenarios, such as co-localisation, where they can localise common objects in a set of images without using object class labels or ground truth locations. The use of STNs allows for more efficient and accurate processing of images by focusing on relevant regions and transforming them to a canonical pose. Overall, STNs provide a flexible and powerful tool for improving the performance of neural networks in various computer vision tasks.