How to modify the model in order to train a ssd with 150*150 inputs?

[MinLiAmoy]

Hi, recently I am considering to train a model and the size of the input image is 150*150. I expect the speed could be faster than ssd300, due to the computation and smaller feature map to generate bbox. How can I modify the code or prototxt? Thanks!

[linux-devil]

check out train.prototxt file , you need to change dimensions there, likewise for test and deploy prototxt

[MinLiAmoy]

@linux-devil thanks for your reply =)

[linux-devil]

You may close this issue

[rmx564354093]

I changed train.prototxt test.prototxt deploy.prototxt resize_param or dimensions from 300 to 150 and then train again but the result of detect is very bad. Do you still need to modify other parameters?

[yokings]

I changed train.prototxt test.prototxt deploy.prototxt resize_param or dimensions from 300 to 150 and then train again but the result of detect is very bad. Do you still need to modify other parameters?

change min_size and max_size in priorbox layer

[soumilchugh]

I am facing the same issue. Updated the min_size and max_size in priorbox layer but the performance is still bad. On the same dataset, the SSD model with image size 300x300 performs very well. Any reasons why on 150x150, it is not training?