Angela Dai is a Ph.D. student at Stanford University, has a Spotlight Talk on CVPR, mainly introducing the ScanNet, a large-scale RGB-D data set with labeling 3D indoor scene reactive information.
Her initial idea is to promote the development of the machine learning algorithm, especially in 3D data. 3D data contains more information, such as the distance between the size and the object. But 3D data is more difficult to get, it is harder to add a label, and now there is not much 3D data.
Angela wants to build an extensible data acquisition framework with Scannet. They first need to collect 3D reconstruction data and then labeled data in a valid manner to collect more data. At present, the team has collected about 1500 RGB-D video sequences, and collects with a depth sensor through the iPad application. The video will then be uploaded to the server and is automatically rebuilt. Then, the video will be given to Amazon Mechanical Turk and will be labeled out.
Large-scale 3D Dataset SCANNET: Let the robots understand the real world
The data label is in a given 3D scene, drawing objects, for example, drawing a chair, a table or computer, thus knowing what is, and location. Each image usually requires 5 people to mark it. The resulting data can be used as a standard reference when making a training task such as an object.
The SCANNet dataset can help train algorithms directly on 3D data. For example, if there is a robot moving in the room, it needs to identify what objects in the room, but not only need to identify an object in the distance, but also determine what this object is.
Angela and team also made a few scenario-understood benchmarks on realistic data. Because the existing large 3D data sets are synthesized, this is very different from the 3D data collected by the real world.
Large-scale 3D Dataset SCANNET: Let the robots understand the real world
Typically, if you train algorithms through the synthesis database, the effect is not too good when the algorithm is used for real data, as the computer does not learn the data characteristics of the real world. There are many noises in reality, it is difficult to observe all of the features of an object. The benchmark test shows that the training effect of the computer in real data is much better than the training effect in the synthesis data. There will be a bigger demand after real data.
Angela has been studying 3D reconstruction, developing a real-time 3D reconstruction system, but she later discovered that it is difficult to apply in practice because lack of semantic understanding of the scene. In a scene, people will want to know which object is in the object, so that there is a virtual assistant or chat robot, help the interaction of the scene. This is why she developed new data sets.
In addition, in addition to the public bidding tasks, they also hope to reconstruct the mission in the prepositories. In addition, there must be a lot of work in semantic understanding. But their current tasks are to solve object identification. 3D Scene Data will have more interesting applications in the future.
Angela is also very interested in combining the data of the real world with the synthetic CAD model. One benefit of this is that the synthetic data is easier to obtain and easy to operate, and if the synthesis data is connected to the real data, the system can be trained on the model, which is easier to migrate to real data.
Of course, the more important task is to give a semantic explanation for 3D data, which facilitates the robot better understand the world.
Original link: https://www.eeboard.com/news/3d-6/
Search for the panel network, pay attention, daily update development board, intelligent hardware, open source hardware, activity and other information can make you master. Recommended attention!
[WeChat scanning picture can be directly paid]
Technology early know:
Never stop millet 5x and miUI9, tomorrow, Millet also releases your own artificial smart product Xiaomi smart speaker?
Heavy pound: End of the "rule" era of operators, will iphone8 will be equipped with E-SIM?
Jia Yueping's LeTTV "script" has arrived in the end, and how is Jia Yue Ting's LeTTTTVTT will change to replacement of the door!
The iPhone8 based on artificial intelligence technology is the winning method of Apple to open the domestic mobile phone manufacturer.
Note! This is the correct posture of using GPS, FM LNA!
Our other product:
