search

rpng / calc

Convolutional Autoencoder for Loop Closure
BSD 3-Clause "New" or "Revised" License
191 stars 45 forks source link

100 landmarks #14

Closed Chatoyant19 closed 5 years ago

Chatoyant19 commented 5 years ago

Hello, thank you for sharing your code. How did you get the 100 landmarks? First get some proposals using BING, and then generate the final detection using a detection? But the results of the final detection are some bounding boxes. I think landmarks should be some points rather than bounding boxes. Thank you very much.

nmerrill67 commented 5 years ago

Typically in SLAM, yes, landmarks are points, but in this paper that we were reproducing with CALC, landmarks are bounding boxes of buildings and other prevalent objects. I would suggest reading the linked paper (cited in our paper as well) to get a better picture of what is going on.

Chatoyant19 commented 5 years ago

Thanks for your response.I train the net from the scratch on the Places dataset(256*256), and I test the model on the campus dataset(your paper).I set the maximum number of iterations to 500k and use iteration 220k(calc_iter_220000.caffemodel) to test the testing dataset as you said in other issues, I also used other iterations to get the caffemodel to test the testing dataset. But my results are much worse than yours. I use the default batch size 256, and a single GPU(GeForce GTX 1080) for training.Can you tell me how to get experimental results similar to yours?

nmerrill67 commented 5 years ago

Please see previous issues that discuss this issue. I used a larger batch size with two GPUs. The default is set to 256 for convenience for people who may not have the 12GB GPUs.

Mingrui-Yu commented 4 years ago

@Chatoyant19 Hello~ Have you trained the net successfully? I'm also interested in retraining the net and I think I'm facing the same problem.