Generation And Optimization Of Synthetic Dataset And Its Application In Rock Analysis

Data annotation for deep learning consumes much time and manpower.Algorithm development efficiency is limited by the difficulty and cost of making datasets.Moreover,dataset made by hand usually contains many wrong labels that may confuse the training process.In contrast,virtual engine-based synthetic dataset generation allows to obtain a large number of effective training samples quickly.The ground truth labels can be extracted automatically by the engine,and the annotation accuracy remains a high level.The difficulty lies in how to construct a distribution-consistent dataset with the given real-world one.Traditional methods focus on designing virtual environments manually.On one hand,this kind of method needs great manual efforts by many experts.On the other hand,it is not changeable along with the target real-world scenes and needs to re-design new virtual environments to match the new scenes.To alleviate the inconvenience caused by manual designing,this paper adopts the learningbased virtual scene generation method,which is able to automatically construct virtual scenes and generate the corresponding synthetic datasets.This method is mainly composed of two parts.First is the construction of parameterized scene generator within virtual engines.By collecting various 3D assets,we are able to construct an universal and content-abundant synthetic dataset simulator,which is used to automatically generate datasets.Simultaneously during the scene construction process,it realizes the editability of the simulator.That means we are able to control the content of simulated datasets by modifying several attributes defined in engine.Second is learning-based attribute optimization method.This paper proposes two kind of methods from the dataset-level and image-level perspective respectively.First,for the dataset-level method,it optimizes the attributes by constraining the distribution similarity between two datasets.The REINFORCE algorithm is employed due to the non-differentiability of the whole system.In light of the huge search space and low optimization efficiency,this paper proposes the scalable discretization-and-relaxation(SDR)method.It discretizes the sampling space through a policy network,and groups the attributes into several parts to further narrow the sampling space.By doing this,local optimum can be efficiently avoided.The feature of this method lies in that it considers attribute correlation during optimization and thus can lead to more reasonable scene structures.Second,for the image-level method,it optimizes the attributes by constraining the content similarity between two images.This method tackles the system’s non-differentiability and makes it to be an end-to-end trainable framework.Specifically,it converts the scene construction process as an attribute searching for virtual engine and image content similarity measurement iterative problem.By training a generative model to fit the input and output of the engine,the gradients can be approximated during the back-propagation process.As a result,gradient descent can be used to optimize the attributes.The feature of this method lies in that it is able to potentially learn the distribution of real-world dataset by fitting each sample.Afterwards,this paper integrates several existing data simulators and proposes an universal synthetic dataset generation platform-Alice.Existing version contains person re-ID,vehicle re-ID and scene segmentation tasks.The main feature of Alice is that the synthetic datasets are fully editable that allows to perform content-level domain adaptation.What’s more,this paper proposes three new real-world datasets accordingly.These datasets are closer to real applications and have more challenges.It enables uniform evaluation of domain adaptation methods and provides data and environment support to better understand the domain gap.Finally,this paper combines the synthetic dataset generation and optimization method with the rock analysis system in tunnel boring machine(TBM).The system mainly employs semantic segmentation technology to extract rock areas in the images and calculates size distribution.For the lack of rock segmentation dataset,this paper leverages virtual engine to simulate synthetic dataset and further optimizes its content through a learning-based method,making it more similar to real-world data.Experimental results demonstrate the optimized dataset can obtain better training accuracy than random one and joint training with real-world dataset can further boost real-world training accuracy.Research of this paper can alleviate the lack of training data and protection of data privacy to a large extend,and helps to accelerate the development of algorithms fundamentally.