# NeRF-Supervision: Learning Dense Object Descriptors from Neural Radiance Fields ## preliminary ### dense descriptors 相對於object dectection跟instance segmentation,使用dense descriptor能對於物體提供更多細部資訊(例如:用來區分物體的不同部位),讓機器人能從指定的部分操作物體(例如:夾住鞋舌撿起球鞋) ![圖片](https://hackmd.io/_uploads/S1LFBxKoT.png) ### NeRF(neural radiance fields) 以神經網路學習場景的隱式表達法,可用於3D重建 訓練神經網路所需的資料: a set of images with known camera poses ## abstract 對於表面會反光的物體深度相機無法得到好的depth map ![圖片](https://hackmd.io/_uploads/BJ5QExYop.png) 進而影響dense descriptor,於是改用NeRF來產生出correspondences ## 流程 用NeRF產生出correspondences,用來訓練Dense Object Nets ![圖片](https://hackmd.io/_uploads/BJwUSOOo6.png) ## method 1. 先訓練NeRF 2. 實驗一:先用NeRF產生depth map,然後再產生出correspondences 發現有些情況下會產生出不正確的correspondences 3. 實驗二:改採用density field 相較於depth map,使用density field能保留完整分布,而depth map則類似於直接取平均 ## result 可以發現不論是採用depth map或是density field,效果都比 COLMAP(一種傳統的Multi-view Stereo方法)及其他好。其中使用density field的結果最佳 ![圖片](https://hackmd.io/_uploads/rJmmyxxh6.png) ## contributions - RGB-sensor only, self-supervised pipeline for learning object-centric dense descriptors - succeed on thin, reflective objects on which depth sensors typically fail - experiments showing that the distribution-of-depths formulation can improve the downstream precision of correspondence models trained on this data, when compared to the single depth alternatives ## Overview Video https://www.youtube.com/watch?v=_zN-wVwPH1s&t=238s