Mmdetection support

[hardikdava]

Description

Support for MMDetection inference result. This method also supports for mmyolo

Get more information on inference datatype from here

Type of change

• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

Docs

• Docs updated? What were the changes:

[SkalskiP]

@hardikdava, will it work with any MMDetection object detector or just selected ones?

[hardikdava]

It will work with all models from mmdetection and mmyolo. These two framework consists all object detection models. Implemented method works with all these models. Later, I will method to work with segmentation also.

[SkalskiP]

Hi @hardikdava 👋🏻! I tried to test this feature, but it failed. Could you take a look at this notebook https://colab.research.google.com/drive/1HHODtyNlhoXG3etx_fjaU4-B6FsUYHhG, and let me know if I did something incorrectly or if we don't cover that case?

[hardikdava]

@SkalskiP can you approve my request to open the notebook?

[SkalskiP]

@hardikdava you should be able to open it now

[hardikdava]

@SkalskiP I will finish this by the end of the day.

[hardikdava]

@SkalskiP Apparently, mmyolo and mmdetection uses two different API for inference. But the output format is same for both. Here is colab notebook for example. I tested for mmdetection and mmyolo model inference instruction is different for mmyolo but inference results successfully converted using sv.Detections.from_mmdetection() method.

# mmdetection method:
from mmdet.apis import DetInferencer
img_path = './demo/demo.jpg'
# Choose to use a config
model_name = 'rtmdet_tiny_8xb32-300e_coco'
# Setup a checkpoint file to load
checkpoint = './checkpoints/rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth'
# Set the device to be used for evaluation
device = 'cuda:0'

# Initialize the DetInferencer
inferencer = DetInferencer(model_name, checkpoint, device)

# Use the detector to do inference
result = inferencer(img_path, out_dir='./output', return_datasample=True)["predictions"][0]

# mmyolo method:
from mmdet.apis import inference_detector, init_detector #mmyolo uses api from mmdet

# Choose to use a config
config = '/content/mmyolo/checkpoints/yolov5_s-v61_syncbn_fast_8xb16-300e_coco.py'
# Setup a checkpoint file to load
checkpoint = '/content/mmyolo/checkpoints/yolov5_s-v61_syncbn_fast_8xb16-300e_coco_20220918_084700-86e02187.pth'

# Set the device to be used for evaluation
device = 'cuda:0'

# Initialize the DetInferencer
model = init_detector(config, checkpoint, device=device, cfg_options={})


# Use the detector to do inference
img_path = './demo/demo.jpg'
result = inference_detector(model, img_path)

[hardikdava]

@SkalskiP Please test it merge it.

[SkalskiP]

Hi @hardikdava 👋🏻 thanks for creating Colab. It was super helpful. Merging.

[hardikdava]

Hi @hardikdava 👋🏻 thanks for creating Colab. It was super helpful. Merging.

yeah, I like your idea about it. It is very easy to test it. I did the same for paddleDet feature.

[SkalskiP]

I think it would make sense to have a demo Google Colab as a requirement when we implement new features. It makes testing super easy.