1 / 7

Object Classification through Deconvolutional Neural Networks Student: Carlos Rubiano

Object Classification through Deconvolutional Neural Networks Student: Carlos Rubiano Mentor: Oliver Nina. Network in Network. Model different from typical CNN that aims to enhance discriminability for local patches in receptive field

axel
Download Presentation

Object Classification through Deconvolutional Neural Networks Student: Carlos Rubiano

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Object Classification through Deconvolutional Neural Networks Student: Carlos Rubiano Mentor: Oliver Nina

  2. Network in Network • Model different from typical CNN that aims to enhance discriminability for local patches in receptive field • Traditional CNN stacks convolutional and pooling layers and then combines layers in fully connected layer at end • Loss of spatial information • Over fitting can occur

  3. Network in Network • NIN’s model: Passed through multilayer perceptron and then through stacked ‘mlp’ convolutional layers that have composing elements which represent micro networks • Essentially NIN’s micro networks seperates input image into segments and classifies them individually whereas traditional CNN will look at different level of features in each convolutional layer and then combine the features at the end • Rather than using fully connected layer, NIN uses a global average pooling layer which natively prevents overfitting • arxiv.org/pdf/1312.4400v3.pdf

  4. Rectifier Linear Unit • Standard way to model neuron’s output: • f(x) = tanh(x) • However very slow • Hinton and Nair: f(x) = max(0, x) • Maxout:

  5. Feature Map Calculations

  6. UT-Interaction Dataset • Dataset with 2 sets of 6 classes with 10 sequences for each class (total of 120 sequences) • Extracted frames from segments • Resized images to fit our network • Calculated the Optical flow Results • Imagenet: 59%, Current system: 64%

  7. What’s Next? • MBH descriptor seperates optical flow into its horizontal and vertical components • Calculated by taking the gradient of the optical flow to get MBHx and MBHy • Implemented NIN, working on testing with it • Combine different ReLU from maxout and NIN

More Related