Terry Taewoong Um proposes deformable convolutional networks. The document discusses introducing learnable offsets to convolutional filters and region of interest pooling layers to allow the networks to spatially transform based on the input data. This helps the networks better adapt to objects of different scales and aspect ratios. Experimental results show deformable convnets achieve state-of-the-art performance on semantic segmentation, object detection and other tasks. Code is available online for others to experiment with these techniques.