Reimplementing the paper: Cascaded Pyramid Network for Multi-Person Pose Estimation
The model uses the COCO dataset which can be downloaded by typing:
chmod +x coco.sh
./coco.sh
The data is going to be saved inside the coco/ folder.
I actually got the wrong idea of the assigment from the beginning and didn't relize until I searched for a pytorch code on Github for reference.
That is the data doesn't need to be cropped from the original. I mean not physically cropped to images but just need to write the program to cut it during the training process. Anyway I did the cutting and save the neccesary information such as keypoints and visual score (0,1,2) to a dataframe for the training and validation data.
python dataprocessing/process_data.py
python train.py
Download the checkpoint here and unzip.
python test.py
The results are shown below, I know that this one is not a perfect one, but if I have more time I think the model will get better.
| Input | Prediction |
|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
Failed cases
| Input | Prediction |
|---|---|
 |
 |
 |
 |
- the model was not finished training yet, then I was not able to test it.
- There was a typo in the code when I created the dataset and I just figured it out on Friday then everything is just like a fresh start. I will keep training and update the weight file and test code as well as the result.
The repo is heavily based on the pytorch version and tensorflow version and the official keras tutorial about keypoint estimation.