--update=add ckpt links

open-mmlab · Jul 12, 2024 · 0f8da98 · 0f8da98
1 parent 755e515
commit 0f8da98
Show file tree

Hide file tree

Showing 4 changed files with 11 additions and 6 deletions.
diff --git a/projects/rtmpose3d/README.md b/projects/rtmpose3d/README.md
@@ -4,18 +4,21 @@
 
 RTMPose3D is a toolkit for real-time 3D pose estimation. It is based on the RTMPose model, which is a 2D pose estimation model that is capable of predicting 2D keypoints and body part associations in real-time. RTMPose3D extends RTMPose by adding a 3D pose estimation branch that can predict 3D keypoints from images directly.
 
+Please refer to our [technical report](https://arxiv.org/pdf/2407.08634) for more details.
+
 ## 🗂️ Model Zoo
 
 | Model                                                      | AP on COCO-Wholebody | MPJPE on H3WB |                                                   Download                                                    |
 | :--------------------------------------------------------- | :------------------: | :-----------: | :-----------------------------------------------------------------------------------------------------------: |
-| [RTMW3D-L](./configs/rtmw3d-l_8xb64_cocktail14-384x288.py) |        0.678         |     0.052     | [ckpt](https://download.openmmlab.com/mmpose/v1/projects/rtmo/rtmo-s_8xb32-600e_coco-640x640-8db55a59_20231211.pth) |
-| [RTMW3D-X](./configs/rtmw3d-x_8xb32_cocktail14-384x288.py) |        0.680         |     0.052     | [ckpt](https://download.openmmlab.com/mmpose/v1/projects/rtmo/rtmo-s_8xb32-600e_coco-640x640-8db55a59_20231211.pth) |
+| [RTMW3D-L](./configs/rtmw3d-l_8xb64_cocktail14-384x288.py) |        0.678         |     0.056     | [ckpt](https://download.openmmlab.com/mmpose/v1/wholebody_3d_keypoint/rtmw3d/rtmw3d-l_8xb64_cocktail14-384x288-794dbc78_20240626.pth) |
+| [RTMW3D-X](./configs/rtmw3d-x_8xb32_cocktail14-384x288.py) |        0.680         |     0.057     | [ckpt](https://download.openmmlab.com/mmpose/v1/wholebody_3d_keypoint/rtmw3d/rtmw3d-x_8xb64_cocktail14-384x288-b0a0eab7_20240626.pth) |
 
 ## Usage
 
 👉🏼 TRY RTMPose3D NOW
 
 ```bash
 cd /path/to/mmpose/projects/rtmpose3d
+export PYTHONPATH=$(pwd):$PYTHONPATH
 python body3d_img2pose_demo.py configs/rtmdet_m_640-8xb32_coco-person.py https://download.openmmlab.com/mmpose/v1/projects/rtmpose/rtmdet_m_8xb32-100e_coco-obj365-person-235e8209.pth configs\rtmw3d-l_8xb64_cocktail14-384x288.py rtmw3d-l_cock14-0d4ad840_20240422.pth --input /path/to/image --output-root /path/to/output
 ```
diff --git a/projects/rtmpose3d/configs/rtmw3d-x_8xb32_cocktail14-384x288.py b/projects/rtmpose3d/configs/rtmw3d-x_8xb32_cocktail14-384x288.py
@@ -128,9 +128,8 @@
                 loss_weight=2.0)
         ],
         decoder=codec),
-    test_cfg=dict(flip_test=False, mode='2d')
-    # test_cfg=dict(flip_test=False)
-)
+    # test_cfg=dict(flip_test=False, mode='2d')
+    test_cfg=dict(flip_test=False))
 
 # base dataset settings
 data_mode = 'topdown'

diff --git a/projects/rtmpose3d/rtmpose3d/pose_estimator.py b/projects/rtmpose3d/rtmpose3d/pose_estimator.py
@@ -67,7 +67,7 @@ def add_pred_to_datasample(self, batch_pred_instances: InstanceList,
                 c = np.array(camera_params['c'])
             else:
                 f = np.array([1145.04940459, 1143.78109572])
-                c = np.array(data_sample.ori_shape)
+                c = np.array(data_sample.ori_shape) / 2
             kpts_pixel = np.concatenate([
                 keypoints_2d,
                 (keypoints_3d[..., 2] + gt_instances.root_z)[..., None]

diff --git a/projects/rtmpose3d/rtmpose3d/simcc_3d_label.py b/projects/rtmpose3d/rtmpose3d/simcc_3d_label.py
@@ -18,6 +18,8 @@ class SimCC3DLabel(BaseKeypointCodec):
     Human Pose Estimation`_ by Li et al (2022) for more details.
     Old name: SimDR
 
+    We generate the SimCC label for 3D keypoint estimation.
+
     Note:
 
         - instance number: N
@@ -93,6 +95,7 @@ def __init__(self,
             root_index, tuple) else [root_index]
 
         # Mean value of the root z-axis of datasets
+        # These values are statistics from the training set
         self.root_z = [5.14388]
         self.z_range = z_range if z_range is not None else 2.1744869