- Creation of a static dataset using blender
- Learn a model on the data
- Use the model to test real scenes
- LDR: Images of 0-255 pixel range, can be displayed on a standard device.
- HDR: Range of luminance that is equivalent to the one experienced by a human eye (upto 105).
- Tone-mapping: The scene with original irradiance(linear space) is mapped to pixel values of (0-255).
- Image from DSLR camera cannot capture much details in very dark or very bright regions in one exposure setting.
- Surveillance or medical applications
- Data set generation
- Learn a network model from the data set
- Run Experiements on the learned model
- Created using Blender seggregating them into 331 train and 75 test images.
- One scene was rendered multiple times in different exposures. The exposures ranged from EV:-5 to EV: +4.
- Render engines are Cycles render and Blender Render.
- The image resolution was kept to 640 X 480.
- 19 Convolution layers, 4 Pooling layers and up-convolution
- 3X3 kernel size, padding of 1, mini batch size 4
- Initial LR of 0.0001, gamma 0.1 and step size 20000
- ReLU activation, Adam Solver
- The experiments were performed on 331 training images and 75 test images.
- Visualization after tone mapping using Gamma Correction
- These experiments were conducted to overcome the problem of having a smaller dataset
- For this we incorporated dropout variations and model reduction
- The HDR Candidates used are obtained by applying an inverse to the camera response function that maps the Low Dynamic Range Images to their corresponding HDR counterparts.
- The HDR Candidates used are obtained by applying an inverse the camera response function that maps the Low Dynamic Range Images to their corresponding HDR counterparts.
- Experiments were performed on the same set of train and test data but we shuffled the stack of LDR images.
- Successfully created a sythetic dataset that can be used for further experimental and research purposes.
- Our network was able to justify on real time images being trained on synthetic data.
- Choice of the network should be taken into consideration as for a smaller dataset a simpler network is more preferable.
- Comparing to architecture from Nima Khademi Kalantari using only an encoder (convolution without de-convolution) can also be an option.
- Real world contains motions
- A real world dataset will contain some motion due to objects moving, hand shaking etc.
- Solving HDR after LDR alignment
- Create a data set that involves motion simulating real time situations.
- Estimating the flow using a FlowNet
- Train the best network from previous findings
- Test generalization on real images with motion