Zhongdao / Towards-Realtime-MOT

Joint Detection and Embedding for fast multi-object tracking
MIT License
2.38k stars 539 forks source link

darknet74.conv trained on what object? #109

Closed PartheshSoni closed 4 years ago

PartheshSoni commented 4 years ago

On what objects(i.e pedestrian only, or general objects) is the darknet model trained on? If I want to train the whole system on some other object, say birds, can I train it on its dataset without changing darknet weights?

WongKinYiu commented 4 years ago

https://github.com/Zhongdao/Towards-Realtime-MOT#pretrained-model-and-baseline-models It is ImageNet pre-trained model.

PartheshSoni commented 4 years ago

https://github.com/Zhongdao/Towards-Realtime-MOT#pretrained-model-and-baseline-models It is ImageNet pre-trained model.

So we don't need to train darknet for objects other than pedestrians?

WongKinYiu commented 4 years ago

This repository joint train detector and embedder simultaneously. image

Zhongdao commented 4 years ago

@PartheshSoni If you want to detect/track other objects than pedestrians, you also need to train with that class of objects as training data. The initial darknet weights are trained on ImageNet dataset for the classification task, it cannot be directly adapted for the detection task.