Yiming-M / CLIP-EBC

The official implementation of the crowd counting model CLIP-EBC.
MIT License
34 stars 6 forks source link

Bird flock Count #3

Closed kushiluv closed 4 months ago

kushiluv commented 5 months ago

Hey, I'm looking for a model which would help me give an estimate on the number of birds in a bird flock. If I use my personal dataset, could you guide me through how I can train this model? Thankyou

Yiming-M commented 4 months ago

Hi @kushiluv,

Our work is currently only aimed at counting people. However, if you are interested into estimating flock sizes, you may try the following steps:

  1. Preprocess your data so that it's organized in the same way as crowd counting datasets. For example:

    Bird Data
    ├── train
    │   ├── images
    │   │   ├── 0001.jpg
    │   │   ├── 0002.jpg
    │   │   ├── ...
    │   ├── labels
    │   │   ├── 0001.npy
    │   │   ├── 0002.npy
    │   │   ├── ...
    ├── val
    │   ├── images
    │   │   ├── 0001.jpg
    │   │   ├── 0002.jpg
    │   │   ├── ...
    │   ├── labels
    │   │   ├── 0001.npy
    │   │   ├── 0002.npy
    │   │   ├── ...

    where the label files are the 2D coordinates of annotated dots.

  2. Add your processed data name in datasets/crowd.py. The ground-truth density map will be automatically generated by Crowd from the provided image and the dot annotations, so don't worry about it.

  3. If you are interested in using CLIP-based models, you may also need to change the text prompt in the format_count function from models/clip/utils.py by replacing "person"/"people" with for example "bird"/"birds".

Please keep me updated with the models' performances in bird counting.

Many thanks, Yiming

969149731 commented 4 months ago

Hello, may I ask what software can be used to annotate data using my own dataset in order to obtain corresponding similar labels The file for 0001.npy, thank you for your answer.

Yiming-M commented 4 months ago

Hello, may I ask what software can be used to annotate data using my own dataset in order to obtain corresponding similar labels The file for 0001.npy, thank you for your answer.

Hi @969149731 ,

I don't have any experience in data annotation. Maybe you could try this one: https://github.com/Elin24/cclabeler, which is used to annotate the NPWU Crowd Counting dataset.

If you are interested, maybe you could also try zero-shot object counting, e.g., https://github.com/songrise/CLIP-Count

Many thanks, Yiming

969149731 commented 4 months ago

Hello, may I ask what software can be used to annotate data using my own dataset in order to obtain corresponding similar labels The file for 0001.npy, thank you for your answer.

Hi @969149731 ,

I don't have any experience in data annotation. Maybe you could try this one: https://github.com/Elin24/cclabeler, which is used to annotate the NPWU Crowd Counting dataset.

If you are interested, maybe you could also try zero-shot object counting, e.g., https://github.com/songrise/CLIP-Count

Many thanks, Yiming

非常感谢你的回答和建议,谢谢!

Yiming-M commented 4 months ago

Hi @kushiluv ,

It's been two weeks that you haven't provided any feedbacks yet. Hence, I presume that the issue has been somewhat resolved, and I am going to close it now. However, please feel free to re-open it if you have further questions.

Many thanks, Yiming