3D Vision back to the start 😆 shape prediction and ingest 3D information 5 data representation depth map RGB + Depth image = RGB-D image 2.5D raw 3D sensor can easily capture depth information another type of depth map is surface normal map, which is a 2D image that represents the surface normal (using a normal vector) at each pixel location all the mentioned maps can be learned with Fully Convolutional Networks (FCN) 🤔 Voxel grid

Reinforcement Learning So far, we have discussed supervised learning, and a little bit unsupervised learning (in UCB-data100 😉) What is Reinforcement Learning at time $t$, env $\rightarrow^{state}$ agent $\rightarrow^{action}$ env $\rightarrow^{reward}$ agent, then env changed, agent learned, then repeated. state can be partial -> noisy reward can be delayed, implicit and sparse -> noisy AND Nondifferentiable 😲 Nonstationary environment, change over time 😎 Generative Adversarial Networks (GANs) somehow is a part of Reinforcement Learning.

Scaling up GNNs 直接load全部nodes又不太可能【naive approach】4090 / A100带不动 neighbor sampling 对hub node的思考 see the paper cluster-GCN advanced Simplified GCN 舍弃了GCN的non-linearity，直接用linear layer 同质性？但是我想知道和glidar的区别？