Cookies help us deliver our services. By using our services, you agree to our use of cookies. Learn more

crowdAI is shutting down - please read our blog post for more information

Mapping Challenge

Building Missing Maps with Machine Learning


Sharing Some tips to Increase the Accuracy, And Report Error of RCNN,

Posted by jestshen92 over 2 years ago

Dear everyone:

I propose to share some tips for increasing the accuracy.

After the last comment of mine, I try to add some data augmentation. Apart from classical data augmentation, i.e., flip, rotation and blur etc. I also add a warped method based on http://faculty.cs.tamu.edu/schaefer/research/mls.pdf. Actually this works and the accuracy increase about 1 percent. Although the training data has sufficient number, the data augmentation still works even the accuracy >93%. I will upload the results images with yours once I have finished all the interations

Unfortunately I still haven’t solved the problem of coco format as I showed in last comment, so I used Mask_RCNN instead.

One other thing I propose to report is the error in Mask_RCNN.

# Gather all JPG files in the test set as small batches
files = glob.glob(os.path.join(IMAGE_DIR, "*.jpg"))
_buffer = []
for _idx, _file in enumerate(files):
    if len(_buffer) == config.IMAGES_PER_GPU * config.GPU_COUNT:
        _buffer = []

This part has error, which leads the iteration in prediction has less number. Please replace ` _buffer = [] after _buffer.append(_file)` and it can work.

Anyway, I greatly appreciate the organizer to share such sufficient number of dataset. Currently I am doing a project about building distribution changes among years, and this dataset reduces the working value of our team for months.

Thanks again!

Best regards,

Shenlu Jiang.


Posted by Joe233  over 2 years ago |  Quote

Hello,Shenlu Jiang. I think you are very good at this and your reslut is very nice. Can you tell me youer email or QQ.I have some problems with this and I wish to consult you on a few questions. Looking forward to hearing from you! Qin Jun.

Posted by Joe233  over 2 years ago |  Quote

My email is qj5657@gmail.com . Looking forward to hearing from you!