One shot detector or two shot detector?

While many comparative studies on two shot object detectors and one shot object detectors wherein most of them claiming two shot detectors like RCNN series offer better accuracy at the cost of slower processing, an attempt was made to draw own conclusions on military objects. Key idea was to establish feasibility.

Objectives:

1. Comparison of algorithms
2. Feasibility on small datasets (less than 100 images)

Results

Shortcomings/ Limitations of the approach

1. Only two algorithms were selected. Faster RCNN - two shot, Yolo v4 - one shot).
2. Better algorithms exist today.
3. Small dataset (but then that was part of the objective).
4. No of classes limited.
5. All algorithms used as is without any parameter tuning.
6. Use of transfer learning.

Conclusion

Given the limitations/ short comings in the approach, one shot algorithm gave better accuracy in addition to better performance.