Fine-tuning MobileNet SSD network -- 4
Budget: €349 – €350 EUR
Needed an experienced deep learning expert for fine-tuning MobileNet SSD network.
On my server I have generated .caffemodel using this network: https://github.com/chuanqi305/MobileNet-SSD.
Five classes of vehicles have to be recognized: car, van, truck, bus and tram.
Input images were taken from 3 locations (3998 photos in total, in training: 20% were taken for test dataset):
Location A + Location B =>11027 cars, 1095 vans, 869 trucks, 1024 buses, 1200 trams (cca. 20% are night scenes)
Location C (accuracy is much worser than loc. A and B) => 915 cars, 142 vans, 166 trucks, 41 buses
All images went into Caffe training process in resolution 800x600 px or 600x800 px (location-C cam is oriented vertically).
TASKS:
1.) Can you optimize network so I don't need to perform more image-collecting for "location-C" (which accuracy is much worse than previous two, regardless of class type).
2.) And more generally for future: optimizations that are applicable for each camera's location where scene is similar to these 3 locations.
3.) Could You detect and correct eventual mistakes in my training process what led to significantly worser accuracy at location C ?
Training log is attached, note that training accuracy is usually not exceeding 75% !
On my server I have generated .caffemodel using this network: https://github.com/chuanqi305/MobileNet-SSD.
Five classes of vehicles have to be recognized: car, van, truck, bus and tram.
Input images were taken from 3 locations (3998 photos in total, in training: 20% were taken for test dataset):
Location A + Location B =>11027 cars, 1095 vans, 869 trucks, 1024 buses, 1200 trams (cca. 20% are night scenes)
Location C (accuracy is much worser than loc. A and B) => 915 cars, 142 vans, 166 trucks, 41 buses
All images went into Caffe training process in resolution 800x600 px or 600x800 px (location-C cam is oriented vertically).
TASKS:
1.) Can you optimize network so I don't need to perform more image-collecting for "location-C" (which accuracy is much worse than previous two, regardless of class type).
2.) And more generally for future: optimizations that are applicable for each camera's location where scene is similar to these 3 locations.
3.) Could You detect and correct eventual mistakes in my training process what led to significantly worser accuracy at location C ?
Training log is attached, note that training accuracy is usually not exceeding 75% !