QiZuo

Besides, I do not find any dataset named paris in README~

I find that there is no cuda used in the reconstruction pipeline. Is it possible to modify the fusion part using cuda?

I will try the simple way before I dive into it~ Many thanks! By the way, I found that the whole fusion part is not precise as ElasticFusion. something dropped?...

I see the multiplicative noise experiments in paper, maybe the kernel after multiplication is non-linear? In real-life, non-linearity can also be caused by image compression or other factors?

好的谢谢，我尝试之后有结论会分享出来的

INFO:root:missing parameters of layers:['_result_parser.params_map_parser.smpl_model.betas', '_result_parser.params_map_parser.smpl_model.faces_tensor', '_result_parser.params_map_parser.smpl_model.v_template', '_result_parser.params_map_parser.smpl_model.shapedirs', '_result_parser.params_map_parser.smpl_model.J_regressor', '_result_parser.params_map_parser.smpl_model.J_regressor_extra9', '_result_parser.params_map_parser.smpl_model.J_regressor_h36m17', '_result_parser.params_map_parser.smpl_model.posedirs', '_result_parser.params_map_parser.smpl_model.parents', '_result_parser.params_map_parser.smpl_model.lbs_weights', '_result_parser.params_map_parser.smpl_model.vertex_joint_selector.extra_joints_idxs'] 我在load teacher模型hrnet32时会报出这个warning，我觉得应该不会影响params_map和center_map层的蒸馏，想问下result_parser在程序中的主要作用是什么？ （回复早上的：smpl_mesh_root_align调整有用，目前mbv3 from scratch 训到132&85的指标）

好的，关于蒸馏我直接用mse计算centermap以及paramsmap的teacher和student的loss，与gt给出的loss一起回传，请问我这样是否合理？(1000倍是我按照大小调的） ``` if self.distillation_learning: distillated_output = self.teacher_model(meta_data, **self.teacher_train_cfg) d_center_loss = 1000*self.distillation_loss(outputs['center_map'], distillated_output['center_map']) d_params_loss = 1000*self.distillation_loss(outputs['params_maps'],distillated_output['params_maps']) # for key,value in distillated_output.items(): # print(key,value.shape) # print(outputs['center_map'].shape,outputs['params_maps'].shape) if not self.model_return_loss: outputs.update(self._calc_loss(outputs)) loss, outputs...

https://user-images.githubusercontent.com/58206232/160743721-09eda6d9-6691-49ee-b4d0-cde57b8f8b74.mp4 您好，请问您知道这样子shape一会变大一会变小是怎么回事嘛？ 这是我的video.yml配置 ` ARGS: tab: 'process_video' gpu: 0 backbone: 'mbv3' model_precision: 'fp32' val_batch_size: 8 nw: 4 model_path: /test/checkpoints/mbv3_cm64_V1_mbv3_on_gpu0,1,2,3_val/23_129.7699_84.7924_mbv3_cm64_V1_mbv3.pkl # '../trained_models/ROMP_ResNet50_V1.pkl' smpl_mesh_root_align: True centermap_conf_thresh: 0.25 make_tracking: True temporal_optimization: True smooth_coeff:...

原始视频在这： https://user-images.githubusercontent.com/58206232/160775349-a247769b-7cfc-4ba1-ab99-beef63df6f99.mp4 frankmocap的han_hand_long.mp4也会有这个问题，视频太大上传不了，这是下载地址 wget https://dl.fbaipublicfiles.com/eft/sampledata_frank.tar

我可能找到抖动的原因了，我尝试把动作导进blender里面，发现下面这种情况： https://user-images.githubusercontent.com/58206232/160832549-dc345e6c-efd8-4011-b2f4-b6d8edaacd5e.mov 然后我在打印了一下cam_trans参数，发现他的z轴实际变化很大 ` outputs = self.net_forward(meta_data, cfg=self.demo_cfg) print(outputs['cam_trans']) ` ` tensor([[0.1460, 0.2900, 1.3764], [0.1620, 0.2769, 1.1936], [0.1804, 0.2648, 1.3556], [0.1628, 0.2759, 1.1918], [0.1716, 0.2663, 1.2661], [0.1736, 0.2654, 1.2757], [0.1844,...