lizhe00 · yu-jingrui · Dec 19, 2025
diff --git a/configs/mvhn_100840/avatar.yaml b/configs/mvhn_100840/avatar.yaml
@@ -0,0 +1,75 @@
+mode: train
+train:
+  dataset: MvRgbDatasetMVHumanNet
+  data:
+    subject_name: 100840
+    data_dir: ../MVHumanNet_plusplus/mvhumannet++_10demo/main/100840
+    frame_range: &id001
+    - 0
+    - 60
+    - 1
+    used_cam_ids:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    - 5
+    - 6
+    - 8
+    - 9
+    - 10
+    - 11
+    - 12
+    - 14
+    - 15
+    load_smpl_pos_map: true
+  pretrained_dir: null
+  net_ckpt_dir: ./results/100840/avatar
+  prev_ckpt: null
+  ckpt_interval:
+    epoch: 10
+    batch: 50000
+  eval_interval: 1000
+  eval_training_ids:
+  - 42
+  - 2
+  eval_testing_ids:
+  - 50
+  - 0
+  eval_img_factor: 1.0
+  lr_init: 0.0005
+  loss_weight:
+    l1: 1.0
+    lpips: 0.1
+    offset: 0.005
+  finetune_color: false
+  batch_size: 1
+  num_workers: 8
+  random_bg_color: true
+test:
+  dataset: MvRgbDatasetMVHumanNet
+  data:
+    data_dir: ../MVHumanNet_plusplus/mvhumannet++_10demo/main/100840
+    frame_range: [0, 60]
+    subject_name: avatarrex_zzr
+  pose_data:
+    data_path: ../thuman4/pose_00.npz
+    frame_range: [2000, 2500]
+#    data_path: Z:/Data/Pose/AMASS/CMU/06/06_13_poses.npz
+#    data_path: Z:/Data/Pose/AMASS/CMU/10/10_05_poses.npz
+#    frame_interval: 4
+  view_setting: free
+  render_view_idx: 13
+  global_orient: true
+  img_scale: 1.0
+  save_mesh: false
+  render_skeleton: false
+  save_tex_map: false
+  save_ply: false
+  n_pca: 20
+  sigma_pca: 2.0
+  prev_ckpt: None
+model:
+  with_viewdirs: true
+  random_style: false
diff --git a/configs/mvhn_100840/template.yaml b/configs/mvhn_100840/template.yaml
@@ -0,0 +1,46 @@
+mode: train
+train:
+  dataset: MvRgbDatasetMVHumanNet
+  data:
+    subject_name: 100840
+    data_dir: ../MVHumanNet_plusplus/mvhumannet++_10demo/main/100840
+    frame_range: &id001
+    - 60
+    - 61
+    - 1
+    used_cam_ids: null
+    mode: nerf
+  net_ckpt_dir: ./results/100840/template
+  prev_ckpt: null
+  save_init_ckpt: false
+  start_epoch: 0
+  end_epoch: 1001
+  ckpt_interval:
+    epoch: 500
+    batch: 1000
+  eval_interval: 10000
+  depth_guided_sampling:
+    flag: false
+    near_sur_dist: 0.05
+    N_ray_samples: 32
+  compute_grad: true
+  lr:
+    network:
+      type: Step
+      initial: 0.0005
+      interval: 40000
+      factor: 0.9
+      min: 8.0e-05
+  loss_weight:
+    color: 1.0
+    mask: 1.0
+    eikonal: 0.1
+  batch_size: 1
+  num_workers: 0
+model:
+  with_hand: true
+  volume_type: diff
+  use_root_finding: true
+  multires: 6
+  use_viewdir: false
+  multires_viewdir: 3
diff --git a/configs/mvhn_100990/avatar.yaml b/configs/mvhn_100990/avatar.yaml
@@ -0,0 +1,75 @@
+mode: train
+train:
+  dataset: MvRgbDatasetMVHumanNet
+  data:
+    subject_name: 100990
+    data_dir: ../MVHumanNet_plusplus/test/100990
+    frame_range: &id001
+    - 0
+    - 60
+    - 1
+    used_cam_ids:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    - 5
+    - 6
+    - 8
+    - 9
+    - 10
+    - 11
+    - 12
+    - 14
+    - 15
+    load_smpl_pos_map: true
+  pretrained_dir: ./results/100990/avatar/pretrained
+  net_ckpt_dir: ./results/100990/avatar
+  prev_ckpt: null
+  ckpt_interval:
+    epoch: 30
+    batch: 50000
+  eval_interval: 1000
+  eval_training_ids:
+  - 1
+  - 1
+  eval_testing_ids:
+  - 9
+  - 14
+  eval_img_factor: 1.0
+  lr_init: 0.0005
+  loss_weight:
+    l1: 1.0
+    lpips: 0.1
+    offset: 0.005
+  finetune_color: false
+  batch_size: 1
+  num_workers: 8
+  random_bg_color: true
+test:
+  dataset: MvRgbDatasetMVHumanNet
+  data:
+    data_dir: ../MVHumanNet_plusplus/test/100990
+    frame_range: [0, 60]
+    subject_name: avatarrex_zzr
+  pose_data:
+    data_path: ../thuman4/pose_00.npz
+    frame_range: [2000, 2500]
+#    data_path: Z:/Data/Pose/AMASS/CMU/06/06_13_poses.npz
+#    data_path: Z:/Data/Pose/AMASS/CMU/10/10_05_poses.npz
+#    frame_interval: 4
+  view_setting: free
+  render_view_idx: 13
+  global_orient: true
+  img_scale: 1.0
+  save_mesh: false
+  render_skeleton: false
+  save_tex_map: false
+  save_ply: false
+  n_pca: 20
+  sigma_pca: 2.0
+  prev_ckpt: None
+model:
+  with_viewdirs: true
+  random_style: false
diff --git a/configs/mvhn_100990/template.yaml b/configs/mvhn_100990/template.yaml
@@ -0,0 +1,46 @@
+mode: train
+train:
+  dataset: MvRgbDatasetMVHumanNet
+  data:
+    subject_name: 100990
+    data_dir: ../MVHumanNet_plusplus/test/100990
+    frame_range: &id001
+    - 60
+    - 61
+    - 1
+    used_cam_ids: null
+    mode: nerf
+  net_ckpt_dir: ./results/100990/template
+  prev_ckpt: null
+  save_init_ckpt: false
+  start_epoch: 0
+  end_epoch: 1001
+  ckpt_interval:
+    epoch: 500
+    batch: 1000
+  eval_interval: 10000
+  depth_guided_sampling:
+    flag: false
+    near_sur_dist: 0.05
+    N_ray_samples: 32
+  compute_grad: true
+  lr:
+    network:
+      type: Step
+      initial: 0.0005
+      interval: 40000
+      factor: 0.9
+      min: 8.0e-05
+  loss_weight:
+    color: 1.0
+    mask: 1.0
+    eikonal: 0.1
+  batch_size: 1
+  num_workers: 0
+model:
+  with_hand: true
+  volume_type: diff
+  use_root_finding: true
+  multires: 6
+  use_viewdir: false
+  multires_viewdir: 3
diff --git a/dataset/dataset_mv_rgb.py b/dataset/dataset_mv_rgb.py
@@ -498,3 +498,50 @@ def load_color_mask_images(self, pose_idx, view_idx):
         color_img = cv.imread(self.data_dir + '/4x/rgbs/%s/%s_rgb%06d.jpg' % (cam_name, cam_name, pose_idx), cv.IMREAD_UNCHANGED)
         mask_img = cv.imread(self.data_dir + '/4x/masks/%s/%s_mask%06d.png' % (cam_name, cam_name, pose_idx), cv.IMREAD_UNCHANGED)
         return color_img, mask_img
+
+## -------------------------------------------------- ##
+## ------- NEW CLASS FOR MVHumanNet DATASET --------- ##
+class MvRgbDatasetMVHumanNet(MvRgbDatasetBase):
+    def __init__(
+        self,
+        data_dir,
+        frame_range = None,
+        used_cam_ids = None,
+        training = True,
+        subject_name = None,
+        load_smpl_pos_map = False,
+        load_smpl_nml_map = False,
+        mode = '3dgs'
+    ):
+        super(MvRgbDatasetMVHumanNet, self).__init__(
+            data_dir,
+            frame_range,
+            used_cam_ids,
+            training,
+            subject_name,
+            load_smpl_pos_map,
+            load_smpl_nml_map,
+            mode
+        )
+        self.img_files = sorted(os.listdir(self.data_dir + '/images/cam_00'))
+        self.img_file_names = [f.rstrip('.jpg') for f in self.img_files]
+
+    def load_cam_data(self):
+        import json
+        cam_data = json.load(open(self.data_dir + '/calibrations.json', 'r'))
+        self.view_num = len(cam_data)
+        self.extr_mats = []
+        self.cam_names = ['cam%02d' % view_idx for view_idx in range(self.view_num)]
+        for view_idx in range(self.view_num):
+            extr_mat = np.identity(4, np.float32)
+            extr_mat[:3, :3] = np.array(cam_data['cam_%02d' % view_idx]['R'], np.float32).reshape(3, 3)
+            extr_mat[:3, 3] = np.array(cam_data['cam_%02d' % view_idx]['T'], np.float32)
+            self.extr_mats.append(extr_mat)
+        self.intr_mats = [np.array(cam_data['cam_%02d' % view_idx]['K'], np.float32).reshape(3, 3) for view_idx in range(self.view_num)]
+        self.img_heights = [cam_data['cam_%02d' % view_idx]['imgSize'][1] for view_idx in range(self.view_num)]
+        self.img_widths = [cam_data['cam_%02d' % view_idx]['imgSize'][0] for view_idx in range(self.view_num)]
+
+    def load_color_mask_images(self, pose_idx, view_idx):
+        color_img = cv.imread(f'{self.data_dir}/images/cam_{view_idx:02d}/{self.img_file_names[pose_idx]}.jpg', cv.IMREAD_UNCHANGED)
+        mask_img = cv.imread(f'{self.data_dir}/masks/cam_{view_idx:02d}/{self.img_file_names[pose_idx]}.png', cv.IMREAD_UNCHANGED)
+        return color_img, mask_img
diff --git a/gen_data/TRAIN_MVHUMANNET++_DATASET.md b/gen_data/TRAIN_MVHUMANNET++_DATASET.md
@@ -0,0 +1,45 @@
+# Train Models Using the MVHumanNet++ Dataset
+
+## MVHumanNet++ Dataset
+Download MVHumanNet++ Dataset from the [official repo](https://github.com/GAP-LAB-CUHK-SZ/MVHumanNet_plusplus).
+
+MVHumanNet++ is very nice for training AnimatableGaussians models. It features:
+- a large number of subjects and outfits 
+- 16 cameras surrounding the subject
+- 60 frames for each subject-outfit-camera and **an extra A-pose frame**
+- camera calibrations, smplx-parameters are accurate and complete
+
+## Preporcessing
+If you want to preprocess one single subject, run
+```bash
+python preprocess_mvhnpp.py --mvhn_subj_path ../../mvhumannet_plusplus/test/100990
+```
+
+You can put multiple subjects in a folder (e.g. `../../mvhumannet_plusplus/test/`) and preprocess them together by running:
+```bash
+python preprocess_mvhnpp.py --mvhn_root ../../mvhumannet_plusplus/test
+```
+
+The script will create `calibrations.json` and `smpl_params.npz` in the subject folder. These files conform to the THuman4 format, as the authors of AnimatableGaussians recommended.
+
+## Training
+
+You can follow [GEN_DATA.md](./GEN_DATA.md) and [README.md](../README.md) to train or directly use the script. From the root of AnimatableGaussians
+```bash
+chmod +x train_mvhnpp.sh
+./train_mvhnpp.sh mvhn_100990
+```
+You must have `template.yaml` and `avatar.yaml` in `./configs/mvhn_100990` for this to work, which we have provided as an example.
+
+### P.S.
+I could not configure the env using the given `requirements.txt`. I am using
+- `python=3.10`
+- `cuda=11.7.1`
+- `pytorch-cuda=2.0.1`
+- `numpy=1.26.4`
+- `pytorch3d=0.7.4` (installed using downloaded pre-compiled file)
+- **`scikit-learn=1.1.3`** (different from `requirements.txt`)
+- **`scikit-image=0.25.2`** (different from `requirements.txt`)
+- **`opencv=4.7.0`** (different from `requirements.txt`)
+- **`setuptools=78.1.0`** (different from `requirements.txt`)
+- `smplx=0.1.28`