Autolabeling 3D Objects With Differentiable Rendering of SDF Shape Priors

Zakharov, Sergey; Kehl, Wadim; Bhargava, Arjun; Gaidon, Adrien

doi:10.1109/cvpr42600.2020.01224

Cited by 97 publications

(93 citation statements)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, in this work, we explore center clicks, located on BEV maps, as weak supervision signals for 3D object detection. 3D Object Annotation: Very few attempts were made to scale up 3D object annotation pipelines [14,15]. [15] lets an annotator place 2D seeds from which to infer 3D segments and centroid parameters, using fully supervised learning.…”

Section: Related Workmentioning

confidence: 99%

“…[15] lets an annotator place 2D seeds from which to infer 3D segments and centroid parameters, using fully supervised learning. [14] suggests a differentiable template matching model with curriculum learning. In addition to different annotation paradigms, model designs and level of human interventions, our model is also unique in its weakly supervised learning strategy and dual-work mode, and achieves stronger performance.…”

Section: Related Workmentioning

confidence: 99%

“…Table 5 reports the evaluation results for our annotation quality on KITTI val set. Two previous annotation methods [14,15] are included. [15] is a fully supervised deep learning annotator, trained with the whole KITTI train set.…”

Section: Performance As An Annotation Toolmentioning

confidence: 99%

“…It only works as an active model, where humans are required to provide object anchor clicks. [14] requires synthetic data for training and relies on MASK-RCNN [41], so it implicitly uses 2D instance segmentation annotations. It works in an automatic mode.…”

Section: Performance As An Annotation Toolmentioning

confidence: 99%

“…Under such a setting, re-trained PointPillars and PointRCNN reach above 96% of their original performance. More essentially, compared with current strongly supervised annotation tools [14,15], our model is able to provide more accurate annotations with much less and weaker supervision, at higher speed ( §5.4). [13]…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Weakly Supervised 3D Object Detection from Lidar Point Cloud

Meng

Yang

Zhou

et al. 2020

Lecture Notes in Computer Science

View full text Add to dashboard Cite

It is laborious to manually label point cloud data for training high-quality 3D object detectors. This work proposes a weakly supervised approach for 3D object detection, only requiring a small set of weakly annotated scenes, associated with a few precisely labeled object instances. This is achieved by a two-stage architecture design. Stage-1 learns to generate cylindrical object proposals under weak supervision, i.e., only the horizontal centers of objects are click-annotated in bird's view scenes. Stage-2 learns to refine the cylindrical proposals to get cuboids and confidence scores, using a few well-labeled instances. Using only 500 weakly annotated scenes and 534 precisely labeled vehicle instances, our method achieves 85−95% the performance of current top-leading, fully supervised detectors (requiring 3, 712 exhaustively and precisely annotated scenes with 15, 654 instances). Moreover, with our elaborately designed network architecture, our trained model can be applied as a 3D object annotator, supporting both automatic and active (human-in-the-loop) working modes. The annotations generated by our model can be used to train 3D object detectors, achieving over 94% of their original performance (with manually labeled training data). Our experiments also show our model's potential in boosting performance when given more training data. Above designs make our approach highly practical and introduce new opportunities for learning 3D object detection at reduced annotation cost.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%