C 7.3 | Spatial Pyramid Pooling | SPPNet Classification | Fast RCNN | Machine learning | EvODN скачать в хорошем качестве

C 7.3 | Spatial Pyramid Pooling | SPPNet Classification | Fast RCNN | Machine learning | EvODN

SPPNet = SPP + Overfeat for Classification To do image classification, the authors of SPPNet, modified the Overfeat Network. They replaced the last Pool layer with Spatial Pyramid Pooling. Unlike Spatial Pyramid Matching, in SPP, they just take the max pool of the image or quadrants instead of calculating HOG/SIFT features and then preparing codebooks using k-means. The major advantage of using SPP is that, you get a fixed length vector at the output irrespective of the input. So, you can now use images of any size as input. Previously, we saw that, in Overfeat, they took a different approach to work with images of different sizes. That is, they implemented the FC layers as convolution operations. Since, we can use images of any size, you can retain the aspect ratio of the image at the input. In the paper they show that Cropping or Warping the image to fixed size (ex. 224x224) reduces accuracy. They also show that, Spatial Pyramid pooling is robust to object deformations. And that might be another reason for improvement in accuracy. Along with this, you can use Image Pyramids to improve accuracy. On Multiple scales of 224, 256, 300, 360, 448, 560 (and averaging the Softmax scores) improves the accuracy by around 1.5%. My question: Is it worth the extra time taken to do the computations? This technique is generic and can be used on any Network Architecture. With this and few other changes, in ILSVRC 2014, they ranked #3 in image classification Lastly, I have another question: Do we really need Multi-level pooling? No experiments done on this aspect in the paper. But, I guess, we will get our answers in one of the future videos. ------------------------ This is a part of the course 'Evolution of Object Detection Networks'. See full playlist here:    • Evolution Of Object Detection Networks   ------------------------ References: Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition - Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun Credits: https://pxhere.com/en/photo/980743 Copyright Disclaimer: Under section 107 of the Copyright Act 1976, allowance is made for “fair use” for purposes such as criticism, comment, news reporting, teaching, scholarship, education and research.