search

ouusan / some-papers

0 stars 0 forks source link

Choosing Appropriate Learning Strategies #22

Open ouusan opened 1 month ago

ouusan commented 1 month ago

1.Learning to Reconstruct 3D Human Pose and Shape via Model-fitting in the Loop（2019） collaborate regression-based (as initial pose) and iterative optimization-based approach. code: No 2.Weakly Supervised 3D Human Pose and Shape Reconstruction with Normalizing Flows(2020)

code：No 3.Coherent Reconstruction of Multiple Humans from a Single Image（2020） interpenetration loss and interpenetration loss(reprojected instance segmenttaion) code：https://github.com/JiangWenPL/multiperson 4.Deep unsupervised 3D human body reconstruction from a sparse set of landmarks(2021)

code:No 5.Skeleton2Mesh: Kinematics Prior Injected Unsupervised Human Mesh Recovery(2021) root joint+10 local 3D rotations termed θDIK+endpoint use feature map encoding silhouette+other points (concat) code:No code poject page: https://sites.google.com/view/skeleton2mesh 6.(BOA)Bilevel Online Adaptation for Out-of-Domain Human Mesh Reconstruction(2021) spatiotemporal bilevel optimization, adaptation for streaming data,random sample in souce data. code:No 7.(DBOA)Out-of-Domain Human Mesh Reconstruction via Dynamic Bilevel Online Adaptation(2021) BOA+exemplar guidance/exemplar retrieval(based on the similarity between the streaming data and the cluster centers)+ Dynamic Update Strategy(key frames will iterate more steps) code:https://github.com/syguan96/DynaBOA 8.Pose2UV: Single-shot Multi-person Mesh Recovery with Deep UV Prior(2022) visible heatmaps and mask to avoid pixel-level ambiguities(reduce the appearance domain gap) UV prior: VAE code: https://github.com/boycehbz/Pose2UV 9.JOTR: 3D Joint Contrastive Learning with Transformers for Occluded Human Mesh Recovery(2023) 3D Joint Contrastive Learning, Fusion Transformer(2D-based regression and 3D-based refinement) code:https://github.com/xljh0520/JOTR 10.ReFit: Recurrent Fitting Network for 3D Human Recovery(2023) Full-frame Adjusted Reprojection(same to CLIFF??),mocap markers, each keypoint-based feature map code:https://github.com/yufu-wang/ReFit 11.Co-Evolution of Pose and Mesh for 3D Human Body Estimation from Video(2023) Adaptive layer normalization，co-evolution attention block. code: https://github.com/kasvii/PMCE 12.Cyclic test-time adaptation on monocular video for 3d human mesh reconstruction(2023) finetune two pretrained model: HMRNet(other pretrained HMR architectures) and MDNet(pretrained with ) code:https://github.com/hygenie1228/CycleAdapt_RELEASE

ouusan commented 1 month ago
  1. Coherency constraints: ...
  2. related: https://arxiv.org/pdf/1812.10766 SMPLR: Deep SMPL reverse for 3D human pose and shape recovery(volumetric heatmap, denoising autoencoder module)(No code)
  3. DIK module relies a minimal set of prior knowledge that defines the underlying kinematic 3D structure
  4. baseline:6-52 Real-time self-adaptive deep stereo https://arxiv.org/abs/1810.05424 6-50 mean-teacher https://arxiv.org/pdf/1703.01780 figure about correlations between the 2D pose re-projection loss and the evaluation metric.

8.visible pose-mask module is inspired by8-42 ,8-43

  1. CoordConv 9-44 https://arxiv.org/abs/1807.03247 Fusion transformer: inspired by 9-24 MDETR https://arxiv.org/pdf/2104.12763 9-34 Moment-DETR https://proceedings.neurips.cc/paper_files/paper/2021/file/62e0973455fd26eb03e91d5741a4a3bb-Paper.pdf
  2. Domain adaptation for 3D human mesh reconstruction Human motion denoising this work is mainly compared with 12-9 12-8 12-43 MDNet architecture is inspired by https://arxiv.org/pdf/2207.01567 Back to MLP: A Simple Baseline for Human Motion Prediction code: https://github.com/dulucas/siMLPe