AttributeError: 'FigureManagerMac' object has no attribute 'window'

[mumu-create]

Your OS

Mac

Your Python version

Python3.13.13 ( python3.11.15 also have this problem)

What is your library version?

pose2sim 0.10.47

Your problem

Hello! I am a first-year graduate student at a sports university in China. First of all, I would like to sincerely thank you for Pose2Sim—it has provided great conceptual inspiration and technical support for my thesis project.I am using a MacBook with the M5 Pro chip. However, I ran into an issue at the very first step: Pose2Sim.calibration(), problem like this:
" **
In [1]: from Pose2Sim import Pose2Sim

In [2]: Pose2Sim.calibration()
Matplotlib is building the font cache; this may take a moment.
Failed to extract font properties from /System/Library/PrivateFrameworks/FontServices.framework/Resources/Reserved/PingFangUI.ttc: FT_Open_Face (ft2font.cpp line 200) failed with error 0x02: unknown file format
Failed to extract font properties from /System/Library/Fonts/Apple Color Emoji.ttc: Non-scalable fonts are not supported
Failed to extract font properties from /System/Library/Fonts/LastResort.otf: tuple indices must be integers or slices, not str
Failed to extract font properties from /System/Library/Fonts/Supplemental/NISC18030.ttf: Non-scalable fonts are not supported
generated new fontManager

---

Camera calibration
On Friday 26. June 2026, 13:23:06
Calibration directory: ~/demo/calibration

Calculating intrinsic parameters...

Camera int_cam01_img:
cam01_01_int.jpg: Corners found.

AttributeError Traceback (most recent call last)
Cell In[2], line 1
----> 1 Pose2Sim.calibration()

File /site-packages/Pose2Sim/Pose2Sim.py:374, in calibration(config)
372 def calibration(config=None):
373 pipeline = Pose2SimPipeline(config)
--> 374 pipeline.calibration()

File /site-packages/Pose2Sim/Pose2Sim.py:212, in Pose2SimPipeline.calibration(self)
209 logging.info("---------------------------------------------------------------------\n")
211 start = time.time()
--> 212 calibrate_cams_all(config_dict)
213 elapsed = time.time() - start
214 logging.info(f'\nCalibration took {elapsed:.2f} seconds.\n')

File /calibration.py:1627, in calibrate_cams_all(config_dict)
1624 calib_fun = calib_mapping[calib_full_type]
1626 # Calibrate
-> 1627 ret, C, S, D, K, R, T = calib_fun(*args_calib_fun)
1629 # Write calibration file
1630 toml_write(calib_output_path, C, S, D, K, R, T)

File /calibration.py:516, in calib_calc_fun(calib_dir, intrinsics_config_dict, extrinsics_config_dict, save_debug_images)
513 # calculate intrinsics otherwise
514 else:
515 logging.info(f'\nCalculating intrinsic parameters...')
--> 516 ret, C, S, D, K, R, T = calibrate_intrinsics(calib_dir, intrinsics_config_dict, save_debug_images=save_debug_images)
517 nb_cams_intrinsics = len(C)
519 # Early write intrinsics in calibration file

File /calibration.py:767, in calibrate_intrinsics(calib_dir, intrinsics_config_dict, save_debug_images, extrinsics_method)
765 # If not satisfied, tries to detect corners
766 plt.close('all')
--> 767 imgp_confirmed, objp_confirmed = findCorners(img_path, intrinsics_corners_nb, objp=objp, show=show_detection_intrinsics)
768 if isinstance(imgp_confirmed, np.ndarray):
769 imgpoints.append(imgp_confirmed)

File /calibration.py:1036, in findCorners(img_path, corner_nb, objp, show)
1034 if var_to_delete in globals():
1035 del globals()[var_to_delete]
-> 1036 imgp_objp_confirmed = imgp_objp_visualizer_clicker(img, imgp=imgp, objp=objp, img_path=img_path)
1037 else:
1038 imgp_objp_confirmed = imgp

File /calibration.py:1381, in imgp_objp_visualizer_clicker(img, imgp, objp, img_path)
1379 ax.imshow(img)
1380 figManager = plt.get_current_fig_manager()
-> 1381 figManager.window.showMaximized()
1382 plt.tight_layout()
1384 # Allow for zoom and pan in image

AttributeError: 'FigureManagerMac' object has no attribute 'window'** "

I also tried using an AI agent to help diagnose the problem, and it seemed to point to an issue with the interactive backend. Forcing the use of TkAgg made the interaction work properly. And I don't have a deep understanding of the code, so I'm not sure if this modification is correct or if it might cause other issues down the line. I would really appreciate it if you could give me some official guidance on this.

Thank you very much for your time and help!

[davidpagnon]

Hi,

Thanks for the appreciation! Sorry, I thought I had answered this a few days ago...

Forcing the use of TkAgg made the interaction work properly.

Are you telling me that you found the fix? In any case, this is very good to know, thanks! I do not expect any problem to come with this modification, but please tell me if there is any! Unfortunately, I do not have a MacOS computer, but I could try to borrow one and see if I can further fix anything.

Another recent issue on MacOS to be aware of: on the latest MacOS 26+ platforms, mps (GPU) support is not available yet, so I fall back to using openvino backend with cpu device. This is relatively slow. I'll change it when it is supported. But in the meantime, can you try openvino backend and gpu device and tell me how it goes? (the option is available in Config.toml)

[mumu-create]

Hello David!

Thank you very much for your reply! ！To be precise, the method "Forcing the use of TkAgg made the interaction work properly" was diagnosed with the help of Kimi, a large Chinese AI model. It currently reproduces the project you provided, "Demo_SinglePerson", very well ( I forgot to mention that after I changed to TkAgg, there is still a minor issue: after completing the calibration of the last extrinsic image, closing the interactive window causes a lag/freeze. Although the extrinsic calibration succeeds, I have to exit IPython and re‑enter it to continue to the next step. This might be caused by my modification ) .

In addition, I will try using the OpenVINO backend with the GPU device in the next few days and let you know how it goes. Finally, thanks again for your help!！

[davidpagnon]

Hi,
I think both Pose2Sim and Sports2D should now work on MacOS :)
Can you test it and tell me how it goes? You will need to install it from source. If it works, I'll make it a release:

  git clone --depth 1 https://github.com/perfanalytics/pose2sim.git
  cd pose2sim
  uv pip install -e .

[mumu-create]

Hi David,
That's fantastic — I tested it as you requested, and the macOS GUI issues are all resolved now! There's no more lag after the external parameter calibration either. Thank you so much for your help!

I also wanted to let you know that, unfortunately, I don't currently have access to an OpenVINO backend or a GPU device on hand, so I can't report on how those perform yet.

In the meantime, I tested the Demo_SinglePerson project. When running Pose2Sim.poseEstimation() with the latest version(with ONNX Runtime 1.27.0 and Protobuf 7.35.1, device=auto，backend=auto), the pose estimation was noticeably slower than the previous version — it used ONNX Runtime + GPU and took 36 seconds to process 4 videos, device=auto，backend=auto. The log is below:

Using model HALPE_26 for Body_with_feet pose estimation in balanced mode.
Inference run only every 4 frames. Inbetween, pose estimation tracks previously detected points. Detections are dropped if their average keypoint likelihood is below 0.5.
Valid MPS installation found: using ONNXRuntime backend with GPU.
Cannot run pose estimation in parallel with display_detection=true. Set it to false for faster pose estimation.
Tracking is performed with sports2d.

Found 4 video files in videos.
Using cpu device with onnxruntime backend.
load ~/.cache/rtmlib/hub/checkpoints/yolox_m_8xb8-300e_humanart-c2c7a14a.onnx with onnxruntime backend
load ~/.cache/rtmlib/hub/checkpoints/rtmpose-m_simcc-body7_pt-body7-halpe26_700e-256x192-4d3e73dd_20230605.onnx with onnxruntime backend
Processing cam01.mp4: 100%|████████████████████████████████████████████████████████████████████████████| 100/100 [00:10<00:00,  9.62it/s]
--> Output video saved to ~/pose2sim/Pose2Sim/Demo_SinglePerson/pose/cam01_pose.mp4.
Processing cam02.mp4: 100%|████████████████████████████████████████████████████████████████████████████| 100/100 [00:10<00:00,  9.66it/s]
--> Output video saved to ~/pose2sim/Pose2Sim/Demo_SinglePerson/pose/cam02_pose.mp4.
Processing cam03.mp4: 100%|████████████████████████████████████████████████████████████████████████████| 100/100 [00:07<00:00, 12.95it/s]
--> Output video saved to ~/pose2sim/Pose2Sim/Demo_SinglePerson/pose/cam03_pose.mp4.
Processing cam04.mp4: 100%|████████████████████████████████████████████████████████████████████████████| 100/100 [00:07<00:00, 12.76it/s]
--> Output video saved to ~/pose2sim/Pose2Sim/Demo_SinglePerson/pose/cam04_pose.mp4.

Pose estimation took 00h00m36s.

---

For comparison, I reinstalled the previous Pose2Sim 0.10.45 version (with ONNX Runtime 1.26.0 and Protobuf 7.35.0, device=auto，backend=auto). Using ONNX Runtime + MPS, it was significantly faster — only 16 seconds for the same 4 videos — and the output preview videos were just as smooth as the originals. Here's that log:

Using model HALPE_26 for Body_with_feet pose estimation in balanced mode.
Inference run only every 4 frames. Inbetween, pose estimation tracks previously detected points. Detections are dropped if their average keypoint likelihood is below 0.5.
Valid MPS installation found: using ONNXRuntime backend with GPU.
Cannot run pose estimation in parallel with display_detection=true. Set it to false for faster pose estimation.
Tracking is performed with sports2d.

Found 4 video files in ./videos.
Using mps device with onnxruntime backend.
2026-07-02 14:17:19.952 python3[85937:8975204] 2026-07-02 14:17:19.952017 [W:onnxruntime:, coreml_execution_provider.cc:113 GetCapability] CoreMLExecutionProvider::GetCapability, number of partitions supported by CoreML: 8 number of nodes in the graph: 415 number of nodes supported by CoreML: 379
2026-07-02 14:17:21.916 python3[85937:8975204] 2026-07-02 14:17:21.916278 [W:onnxruntime:, session_state.cc:1367 VerifyEachNodeIsAssignedToAnEp] Some nodes were not assigned to the preferred execution providers which may or may not have an negative impact on performance. e.g. ORT explicitly assigns shape related ops to CPU to improve perf.
2026-07-02 14:17:21.916 python3[85937:8975204] 2026-07-02 14:17:21.916304 [W:onnxruntime:, session_state.cc:1369 VerifyEachNodeIsAssignedToAnEp] Rerunning with verbose output on a non-minimal build will show node assignments.
load ~/.cache/rtmlib/hub/checkpoints/yolox_m_8xb8-300e_humanart-c2c7a14a.onnx with onnxruntime backend
2026-07-02 14:17:22.868 python3[85937:8975204] 2026-07-02 14:17:22.868270 [W:onnxruntime:, session_state.cc:1367 VerifyEachNodeIsAssignedToAnEp] Some nodes were not assigned to the preferred execution providers which may or may not have an negative impact on performance. e.g. ORT explicitly assigns shape related ops to CPU to improve perf.
2026-07-02 14:17:22.868 python3[85937:8975204] 2026-07-02 14:17:22.868297 [W:onnxruntime:, session_state.cc:1369 VerifyEachNodeIsAssignedToAnEp] Rerunning with verbose output on a non-minimal build will show node assignments.
load ~/.cache/rtmlib/hub/checkpoints/rtmpose-m_simcc-body7_pt-body7-halpe26_700e-256x192-4d3e73dd_20230605.onnx with onnxruntime backend
Processing cam01.mp4: 100%|██████████████████████████████████████████████████████████| 100/100 [00:03<00:00, 28.00it/s]
--> Output video saved to ~/Desktop/AllProject/AcademicDissertation/pose2sim01045/demo/pose/cam01_pose.mp4.
Processing cam02.mp4: 100%|██████████████████████████████████████████████████████████| 100/100 [00:03<00:00, 28.87it/s]
--> Output video saved to ~/Desktop/AllProject/AcademicDissertation/pose2sim01045/demo/pose/cam02_pose.mp4.
Processing cam03.mp4: 100%|██████████████████████████████████████████████████████████| 100/100 [00:03<00:00, 31.02it/s]
--> Output video saved to ~/Desktop/AllProject/AcademicDissertation/pose2sim01045/demo/pose/cam03_pose.mp4.
Processing cam04.mp4: 100%|██████████████████████████████████████████████████████████| 100/100 [00:03<00:00, 30.94it/s]
--> Output video saved to ~/Desktop/AllProject/AcademicDissertation/pose2sim01045/demo/pose/cam04_pose.mp4.

Pose estimation took 00h00m16s.

Thanks again for all your hard work—it's been a huge help!！

[davidpagnon]

That was fast, great to know!

Can you try device='onnxruntime' and backend='mps'?
If it does not work or is not noticeably faster, can you try to uninstall onnxruntime 1.27.0, install onnxruntime 1.26.0?

Can you also tell me what your macOS version is (essentially, > or < to MacOS 26)?

The reason I'm asking you all that is that 'mps' device is not always supported anymore. But I'm not sure whether it is a matter of MacOS version or of onnxruntime version.

[mumu-create]

Absolutely, David! my macOS version is 26.5.1.

I tested three configurations with onnxruntime 1.27.0:

1、device='onnxruntime' and backend='mps' – Pose estimation failed; no skeleton or keypoints appeared in the video, and the video playback was very slow. Partial logs are as follows:
Using mps device with onnxruntime backend. 2026-07-02 22:43:31.067 python3[20369:9317721] 2026-07-02 22:43:31.067936 [W:onnxruntime:, coreml_execution_provider.cc:137 GetCapability] CoreMLExecutionProvider::GetCapability, number of partitions supported by CoreML: 7 number of nodes in the graph: 415 number of nodes supported by CoreML: 383

2、device='onnxruntime' and backend='cpu' – Similar to auto; relatively slow (~35 seconds), with noticeable lag during recognition and increased memory pressure (the indicator turned from green to yellow).

3、device='openvino' and backend='cpu' – This combination was somewhat faster (~25 seconds) and the video playback was smoother during recognition. Although still noticeably slow, it performed better than ‘onnxruntime/cpu’ and the ‘auto/auto’ setting, and there was no memory pressure (remained green).

Then I downgraded to onnxruntime 1.26.0 and used device='onnxruntime' with backend='mps'. This time it was fast (16 seconds), consistent with my previous experience, However, after the test finishes, it gets stuck on the last video. As soon as I run Pose2Sim.synchronization(), it works properly. For your reference, here are the logs from that run:
Using mps device with onnxruntime backend. 2026-07-02 23:11:38.698 python3[27449:9367268] 2026-07-02 23:11:38.698029 [W:onnxruntime:, coreml_execution_provider.cc:113 GetCapability] CoreMLExecutionProvider::GetCapability, number of partitions supported by CoreML: 8 number of nodes in the graph: 415 number of nodes supported by CoreML: 379

I hope this helps. Please let me know if you need any additional details or further tests.

[mumu-create]

Oh, by the way, I'd like to add that with onnxruntime 1.26.0, using device='auto' with backend='auto' gave similar performance to (device='openvino' and backend='cpu' on onnxruntime 1.27.0), taking 22 seconds. Partial logs are as follows:

Using model HALPE_26 for Body_with_feet pose estimation in balanced mode.
Inference run only every 4 frames. Inbetween, pose estimation tracks previously detected points. Detections are dropped if their average keypoint likelihood is below 0.5.
Valid MPS installation found: using ONNXRuntime backend with GPU.
Cannot run pose estimation in parallel with display_detection=true. Set it to false for faster pose estimation.Tracking is performed with sports2d.

[davidpagnon]

Okay, thank you very much!

So onnxruntime 1.26.0 fixes the issue when onnxruntime and mps are used, this is very good to know.

It makes sense that auto gives you the same processing time as onnxruntime and cpu, because this is what I set it to do temporarily, just to make sure it passes the tests 😁
https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/poseEstimation.py#L315

I'll change this like back to what it used to be, pin onnxruntime<1.27, and it should work! There is still the problem of Pose2sim being stuck in the last video, but to be honest, since I gave the macBook back, I have no idea what the problem is 😥

[mumu-create]

David, I think this is a minor issue for you, so it's OK，and thank you so much! But the above was just my personal attempt. I think if you have time, you could verify it yourself. I hope my tests haven't misled you, because I did run them twice and that's indeed what I found. I'll test it again tomorrow.

[davidpagnon]

Thanks for your testing and your kind words!
My fix is not perfect, but it is slightly better than the current situation: if the detected onnxruntime version is >v1.26.0, I give the user a warning message that recommends downgrading. Until it is done, auto switches to onnxruntime with cpu. The reason I do not fall back to openvino is that it does not work on macOS for RTMO models.

[mumu-create]

You're very welcome!
I will be carrying out my pilot experiment over the next few weeks. I hope everything goes well, and I wish you all the best too!

[davidpagnon]

Okay, I think I finally:

• Fixed the display issue
• Did my best on the device / backend one. With 'auto' on MacOS, it checks which ONNXRuntime version is installed. If it is <=1.26.0, it uses onnxruntime and mps. If not, it switches to onnxruntime and cpu, and a warning message recommends downgrading.

Hopefully, this is now good enough! But if it is not, please tell me :)
(no rush, though)

(Also, are you on LinkedIn? Once it reliably works, I'd like to post there and give you and the other guy a shoutout for helping me test things out)

[mumu-create]

I'm so pleasantly surprised by this, but unfortunately I don't have LinkedIn.

If this update actually makes everything run stably, that would be awesome. You really don't need to give me credit – I don't think I helped that much. If you'd like to refer to me somehow, just say "a student from China" , it would be my honor!