Tri-camera module with IMU and dual/triple Cortex M7
Software
└── DeepLearning
├── HomographyNetUnsup
| └── *
├── RealData
| └── *
├── SimilarityNet
| └── *
├── SuperPoint
| └── *
└── SpeedTests
└── *
HomographyNetUnsuphas the depracated code from EVDodgeNet.RealDatahas the code for evaluation on data from simulation and real world.SimilarityNethas the code for training and evaluation for all the scenarios presented in the paper.SpeedTestshas the code for testing speed of various networks.
To train use TrainSimilarityNet.py. The following command line flags can be used.
--BasePath: Base path from where images are loaded, eg., /home/nitin/Datasets/MSCOCO/train2014Processed.--NumEpochs: Number of epochs the training will be done for--DivTrain: Factor to reduce Train data by per epoch, used for debugging only or for super large datasets.--MiniBatchSize: Size of the MiniBatch to use.--LoadCheckPoint: Load Model from latest Checkpoint from CheckPointPath?--RemoveLogs: Delete log Files from ./Logs? (BUGGY, DON'T USE).--LossFuncName: Choice of Loss functions, choose from SL2 (Supervised L2 loss), PhotoL1 (Photometric L1), PhotoChab (Photometric Chabonnier), SSIM (SSIM Loss from GeoNet), SSIMTF (TF's Implementation of GeoNet's SSIM loss). Any loss can be suffixed with HP or SP to compute loss on Highpassed image or SuperPoint cornerness respectively.--RegFuncName: Choice of regularization function, choose from None, SP (Cornerness from SuperPoint) or HP (Highpassed image). Total loss is computed as LossFunc + sum(Alpha_i * RegFunc_i)--NetworkType: Choice of Network type, choose from Small, Large.--NetworkName: Name of network file, eg., Network.VanillaNet. Change the lineVN = Net.xxxinTrainOperationfunction to use the correct class name of the network.--CheckPointPath: Path to save checkpoints.--LogsPath: Path to save Logs.--GPUDevice: What GPU do you want to use? -1 for CPU.--DataAug: Depracated, DO NOT USE. Instead Augment data before and create a dataset.--LR: Learning Rate.--InitNeurons: Number of Neurons in the first layer--Input: Input, choose from I: RGB Images, G: Grayscale Images, HP: HP Grayscale Images, SP: Cornerness from SuperPoint.
To run the code use CreateNetwork.py. The following command line flags can be used.
--CheckPointPath: Path to save Model, eg., /home/nitin/PRGEye/CheckPoints/SpeedTests--ModelPrefix: Prefix for model name, Useful if you are running for different variants of the model.--GPUDevice: What GPU do you want to use? -1 for CPU--MiniBatchSize: Size of the MiniBatch to use--NetworkName: Name of network file, eg., Network.VanillaNet--TestMultipleMiniBatch: Used for Testing only, Test for running at different batch sizes. If active MiniBatchSize represents max value.--NumTest: Number of times code will run to take Avg. time.--Mode: Train creates the Model. Test runs the speed tests. Don't use TrainTest mode.--MiniBatchMin: Used for Testing only, is active when TestMultipleMiniBatch is used and MiniBatchMin represents min value.--MiniBatchSizeIncrement: Used for Testing only, is active when TestMultipleMiniBatch is used and MiniBatchSize represents max value. The code is tested for these increment steps.--ForceBatchSize1: Used for Testing only, is active when TestMultipleMiniBatch is used and MiniBatchSize represents max value. Force runs for BatchSize of 1.--TFLite: Convert to TFLite quantized to type --TFLiteQuant?--TFLiteOpt: TFLite Optimization, Choose from Default, Size and Latency.--TFLiteQuant: TFLite Quantization. Choose from Float32, Float16, Int64, Int32, Int8, UInt8. Default: Float32--EdgeTPU: TFLite For EdgeTPU. This works in addition to TFLite conversion.--InitNeurons: Number of Neurons in the first layer
Additional Notes:
- Use TF 1.15 in Python 2.7 for Train mode in
SpeedTests. - Use TF 1.14 in Python 3.6.9 for Test mode in
SpeedTests.