1. Use Case : Learned Depth MAPS Compression
1.1. Compression Architecture
1.1.1. Encoder: WellDCFNet (without Softmax layer).
1.1.2. Decoder: Reverse of encoder layers.
1.1.3. Entropy model: Hyperprior
1.2. Training
1.2.1. Dataset: DepTex_10000
1.2.2. Batch size: 8, Epochs: 100, Learning rate: 10^-4
1.2.3. Loss function: Rate-Distortion (R + λD
1.3. Preliminary Results
1.3.1. Competitive Rate/PSNR performance compared to 3D-HEVC, especially at middle and high bitrates
2. Conclusion
2.1. Contributions
2.1.1. WellDCFNet model for depth map feature extraction
2.1.2. Combines wedgelet filters and Learnt Deep Correlation Features (LDCF)
2.2. Future Work
2.2.1. Improve entropy model for better rate estimation
2.2.2. Develop a variable-rate compression model to reduce computational requirements
3. Context
3.1. Emergence of depth sensors
3.2. Applications
3.2.1. face recognition
3.2.2. Gesture recognition
3.2.3. Segmentation
4. Objective
4.1. Automatic feature extraction model based on an optimized CNN
4.2. Texture Distinction
4.3. Deep Learning
5. Feature Extraction Approaches
5.1. Traditional Approaches
5.1.1. Fourier Transform
5.1.2. Radon Transform
5.1.3. Edge detection filters
5.1.3.1. Canny
5.1.3.2. Sobel
5.1.3.3. Prewitt
5.1.4. HOG , SURF , SIFT , Hough Lines
5.2. Deep Learning Approaches
5.2.1. CNNs Trained from Scratch
5.2.1.1. Regular CNN
5.2.1.2. FCNN
5.2.1.3. CNN-PPD
5.2.2. Pre-trained CNNs
5.2.2.1. DenseNet
5.2.2.2. Xception
5.2.2.3. VGG-19
6. WellDCFNet Model
6.1. Architecture
6.1.1. First convolutional layer with pre-defined wedgelet filters.
6.1.2. Pre-trained VGG-19 network.
6.1.3. Learnt Deep Correlation Features (LDCF) for style classification
6.2. Wedgelet Filters
6.2.1. Partition of an image block into two regions separated by a straight line
6.2.2. Used in 3D-HEVC for depth map modeling
6.3. Feature Maps Correlation
6.3.1. Learnt intra-layer and inter-layer correlations
6.3.2. Neural network to extract global features distinguishing depth maps from texture images
7. Experiments
7.1. Dataset Preparation
7.1.1. DepTex_10000*: 10,000 images (5,000 texture and 5,000 depth maps)
7.1.2. Convert color texture images to grayscale
7.2. Comparison with Other Architectures
7.2.1. WellDCFNet vs Regular CNN, FCNN, CNN-PPD, DenseNet, Xception, VGG-19