Fully Convolutional Networks for Semantic Segmentation: ConvNets are powerful models not only for image classification, but also for semantic segmentation. This paper show that a fully convolutional network (FCN), trained end-to-end, exceeds the state-of-the-art on semantic segmentation without further machinery. An input image is first downsampled by convolutional layers and then upsampled to same size as input to get the predicated segmenatation map.
star64_v2.ipynb : I extract false positive images for Conv2D models with different filter size. And I take a look at what visual patterns models learned by creating input images that maximize the activation of specific filters.
Figure 1: The confusion matrixes for convolutional models with three different kernel size on test set. They all got accuracy~0.99 on training set, but on test set, larger filter size, lower score on test set.
Figure 2.: The visual patterns corresponding to the first convolutinal layer. Small filter size learns more details on small textures; but less over large scale.
I visualize what patterns Conv2D models with different kernel size learned. For 64*64 light echo images, kernel_size=3 got best result. Larger filter size, lower score on validation and test set. Larger filter size capture more information over large scale; but less in small scale. Larger filter size also means more parameters, therefore easier to overfitting on training set, and thus lower score on test set.
Learn to run Faster RCNN, network structure, input data format and so on. (Delivery 0702)
Build a dataset from ALTAS images for YOLO and RCNN.
Test Conv2D model with dataset at different noise levels.
Learn how to evaulate objection detection model, ROC curve, mAP...
Improve noise model of ALTAS images.