thanks for sharing. I run the code to test my custom video, but find that the result is not so good. the arms will be wrong even for clear image. what the reason do you think it is ?

I used NMR to render the image, but I found that the images were reversed. The images are shown below.

And I found that in function convert_camT_to_proj_mat, you rotate around Y axis and negate the y translation. After I commented on the lines to negate y rotation and y translation, the results seem correct.

https://github.com/saidwivedi/DSR/blob/7d55d173192a44ec567864befd2fb20e993098eb/dsr/semantic_rendering/data_utils.py#L151-L162

Why do you do that extra operation? Did I make some mistakes or get something wrong? Thanks.

Hello. Thanks for your great work! I want to train DSR, but I found that the description of DSR was not quite clear as shown below.

dataset_folders # Contains all datasets
├── coco
│   ├── grph_sequences # Contains psuedo-GT Graphonomy Segmentation
│   └── ...
└── h36m
│   ├── grph_sequences # Contains psuedo-GT Graphonomy Segmentation
│   └── ...
└── ... # Other datasets

I worry that the time spent preparing the data may be too long. Would you like to share the datasets on Google Drive or offer some scripts to process the raw datasets to the specified form of data? Thanks!

Thank you for the cool project. I'm able to run the openpose demo. I want to try using a manually specified bbox. Ignoring the fact that the bbox here isn't very tight, it fails execute on the demo image:

$ python demo.py --checkpoint dsr_data/dsr_w3DPW_checkpoint.pt --img examples/COCO_val_0544.jpg --bbox [0,0,640,427]
2021-10-16 21:01:17.524 | INFO     | dsr.models.hmr:load_pretrained:40 - Loading pretrained weights from dsr_data/dsr_w3DPW_checkpoint.pt
Traceback (most recent call last):
  File "demo.py", line 115, in <module>
    img = process_image(args.img, args.bbox, args.openpose, input_res=constants.IMG_RES)
  File "demo.py", line 84, in process_image
    center, scale = bbox_from_json(bbox_file)
  File "demo.py", line 60, in bbox_from_json
    with open(bbox_file, 'r') as f:
FileNotFoundError: [Errno 2] No such file or directory: '[0,0,640,427]'

Removing the [ ] brackets does not help.

Hi,

I'm trying to download the data, but I'm having the following issue:

$ source prepare_data.sh
--2021-10-16 20:00:52--  https://www.dropbox.com/s/k7bc579qhyxe6ed/dsr_data.zip
Resolving www.dropbox.com (www.dropbox.com)... 2620:100:6019:18::a27d:412, 162.125.4.18
Connecting to www.dropbox.com (www.dropbox.com)|2620:100:6019:18::a27d:412|:443... connected.
HTTP request sent, awaiting response... 301 Moved Permanently
Location: /s/raw/k7bc579qhyxe6ed/dsr_data.zip [following]
--2021-10-16 20:00:52--  https://www.dropbox.com/s/raw/k7bc579qhyxe6ed/dsr_data.zip
Reusing existing connection to [www.dropbox.com]:443.
HTTP request sent, awaiting response... 404 Not Found
2021-10-16 20:00:52 ERROR 404: Not Found.

unzip:  cannot find or open dsr_data.zip, dsr_data.zip.zip or dsr_data.zip.ZIP.
rm: cannot remove 'dsr_data.zip': No such file or directory

Has something happened to the dataset?

Thanks for your great work! I am trying to train DSR recently, but I got 20k+ COCO samples, 90k+ MPI-INF-3DHP samples and 310k+ Human3.6M samples, which results in nearly 24 hours per epoch on 2080ti. That's quite tough for me. I am wondering how can I train such big datasets. It would help me a lot if you could share more training or sampling details.