The Weakly Supervised Learning Method Of Single-image Depth Estimation

Artificial intelligence,a topic that has been mentioned for many times in the two sessions this year,has begun to quietly change our lives.It attempts to make the computer learn,speak,read and write in the fields of image,speech,and natural speech through human thinking,thus combining the advantages of humans and computers to create a new agent that serves humans.The mature application of deep learning algorithm has become one of the opportunities for the rapid development of artificial intelligence.Since the Alex Net model was launched on the Image Net competition in 2012,deep learning has become the focus of major companies and scientific research institutions in the research and application of images,speech,and natural language processing.In particular,convolutional neural networks have caught up with humans in image recognition tasks,making artificial intelligence more than just a science fiction movie.Smart security,smart homes,smart medical care,and the shadow of drones everywhere can be seen.Really improve human life.Firstly,this paper elaborates and analyzes the difficulties of deep learning in single image depth estimation,including the scene ambiguity due to single image,the cost of data collection in the course of learning and training,and the mobile phone running on the mobile terminal.It is difficult to achieve real-time problems.And briefed and analyzed some of the more popular solutions.Followed by the difficulties encountered in the depth estimation put forward some of their own methods and recommendations.First,for the ambiguity of the scene,I chose to modify it based on the original loss function,so that the two data that have different depth values,but the depth value becomes a scalar multiple relationship have the same loss,which introduces scale invariability,eased the scene ambiguity problems caused by a single image.At the same time,since the mean square loss function of the regression task has high sensitivity to outliers,Huber Loss Function is introduced to alleviate this problem,making the model more robust;second,the data samples needed for depth estimation need to be pixel by pixel Points are labeled,so use weakly supervised learning methods for training.For each training set,we regularly and randomly select several points and mark the relative position relationship as the label of our training.We modify the loss function to perform fine-tuning training of the model on the pre-training model trained in the first step.The final output result is still the absolute depth value.Third,the application of the deep learning model at the mobile end has become a mainstream trend.The most important thing to concern is the efficiency of training and forecasting.Try to make the task give results in real time.The model needs to find a balance between accuracy and speed,so I use the ICNet network structure to use multi-resolution image information for multi-channel convolution operations.High-resolution images go through shallow convolution networks,and low-resolution images go through deep layers.The convolutional network not only ensures the semantic information and edge details of the image but also reduces the training and prediction time of the model.Finally,we performed pre-training,fine-tuning,and experimental testing on the standard absolute depth RGB-D data set and the standard relative depth data set.At the same time,two evaluation criteria are given as comparison with other methods.The final experimental results prove that our method can effectively alleviate the problems encountered by the current depth estimation and improve the accuracy and efficiency.The weak supervision training method here can also be used as a general training method,which can effectively reduce the cost of collecting data sets.It is worth learning from.