Customized implementation of the Stereonet guided hierarchical refinement for real-time edge-aware depth prediction
STRUCTURAL ANALYSIS OF DIPS - STEREONETS. Stereonet overlay for low angle dip (reproduce on clear film) The data for each problem are plotted on the tracing paper or overlay, and the stereonet is rotated to suit the differing orientations met with in each case. Although it is usually more convenient to lay the stereonet down and keep it.
Attention: Not accomplished yet
- Stereo Matching is not my main research field, and this repo is created for a homework. So maybe it's not very completed, but I have tried to make it perfect. If you need a better version, please refer to https://github.com/meteorshowers/StereoNet.
- The approach of computing the cost volume in the StereoNet paper is subtracting the padding image and the other image. Here I changed it to concatenate the two images. If you want to change it to the paper's way, just set it when you initialize the net.
- Only training and testing on the KITTI 2015 train dataset is not enough, the best performance has achieved 74.5% (pixels with error smaller than 1). After pretraining on SceneFlow and finetune on KITTI15, the acc achieves 90.054%, not as good as the acc in paper. I have try hard to achieve the accuracy in paper, but still can't. Maybe some details are wrong.
Experiment Results till now
- train and test on SceneFlow datasets:
- epoch 22 total training loss = 3.956
- average test EPE = 3.496
- different finetuning on kitti 15 and result
- 300 epochs, max 3 pixel error rate = 80.893 on kitti val
- 300 epochs, max 3 pixel error rate = 83.527 on kitti val
- 300 epochs, max 3 pixel error rate = 90.054 on kitti val
- 2000 epochs, max 3 pixel error rate = 93.680 on kitti val, after 4.98 hours finetune
Pre-requirement
- Pytorch 1.0.0
- CUDA Toolkit 10
- numpy
Datasets:
- Pretrain: SceneFlow
![Stereonet 8 Stereonet 8](https://www.geosupplies.co.uk/acatalog/stereonet.png)
You can use the anaconda virtual environment to quick start
Install Anaconda
Please reference to Ubuntu系统下Anaconda使用方法总结 for more information about conda installation.
Create Virtual Environment according to my environment index
Training and Test
Switch to the correct python environment
Start training and test
Pretrain on SceneFlow dataset
Finetune on KITTI 2015
Coding Reference
Stereonet Mobile is designed to help you collect, visualize, and analyze geological orientation data on the go. It is a companion to my Stereonet program for desktop computers, but it can also be used to measure orientations on the outcrop. The app has the following features:
Home Screen
Sighting View
Plot Configuration
- Provide automatic data entry as a geologist's compass by reading the iOS device orientation and converting it to strike and dip of planes and/or trend and plunge of lines. It is the only geology compass app that can record the strike and dip of planes using the sighting method and the device camera.
- Provide manual data entry via either tapping and dragging on the stereonet or by typing values into text fields
- Plot one group of lines and one group of planes, though individual observations can be tagged and filtered as the user wishes.
- Carry out basic calculations such as: rotations about any axis, planes from poles or poles from planes, measure angles between lines and between planes, and basic slope stability calculations
- Plot mean vectors, cylindrical best fits, rose diagrams, and contour lines.
- View the plots from any orientation, not just looking straight down into the lower hemisphere
- Stamp each entry with time date and, if enabled by the user, location. See the location of a datum in the iOS Maps program
- Show realtime changes to the stereonet projection to help students understand the geometry of the device better
- Provide basic input and output of data using standard iOS conventions
- Work on both iPhones or iPads, both in portrait mode only. The program will work on iPods or iPads lacking a cellular data connection but those devices have fewer sensors and thus cannot be used in automatic data collection mode.
- Integrated with the StraboSpot online structure and tectonics database. Make your measurements in Stereonet Mobile and upload them directly to your StraboSpot account.
The desktop version of Stereonet has been updated so that it can read and write Stereonet Mobile files, which are simple tab separated text files.
Use it at your own risk. In particular, recall that your iOS device is very sensitive to magnetic fields and the compass bearings can easily be off by 10-20°! This program should not be used as a sole source of navigation data. The author assumes no responsibility for incorrect results, unanticipated behaviors, loss resulting from results provided by the program, or any other anticipated or unanticipated event related to the program.