Is there a General purpose object detection model in NGC?

[music1913]

Hi,
Just like the PascalVOC dataset which covered common objects (20 classes), is there any existing models in NGC, can use directly to perform generic object detection by DeepStream? or we have to train by self?

[mchi]

Plese check TAO Toolkit Integration with DeepStream — DeepStream 6.1.1 Release documentation

[music1913]

thanks, i looked the object detection related deepstream samples and ngc models, seems they are
for purpose build , like trafficcam and dashcam models, they only for 4 classes of objects, and all focus on cars.

I want a model to detect common life objects, like the 20 classes in pascal voc dataset(ms coco is good as well). is there any one?

[mchi]

@music1913

purpose build

what do you mean purpose build? Did you find YoloV3, SSD, etc in the page?

[music1913]

thanks.
these are the models i refered, they’re out-of-box to use, but object classes is limited(4):

(I’m quite new for this) those YoloV3, SSD and etc are all pre-trained on very limited object classses as I see their lables.txt, correct?

[mchi]

you can retrain either one with your dataset to support more classes

[music1913]

just curious why there’s no many-class detection models, thanks anyway.

[Morganh]

@music1913 ,
All the pretrained models released by TAO are trained on NVIDIA internal dataset.

For public dataset, for example, VOC or COCO or KITTI, we did not release models. But we release some guides to tell end user how to train with the public dataset. For example, https://developer.nvidia.com/blog/preparing-state-of-the-art-models-for-classification-and-object-detection-with-tlt/