search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.41k stars 11.66k forks source link

objects are completely detected in one half of the image, but not detected at all on the other half #536

Open baharna2 opened 6 years ago

baharna2 commented 6 years ago

out_newmarket out_newmarket2 out_newmarket3 out_newmarket4 I used mask_RCNN to detect the cars in the parking lot, I'm not sure why it worked pretty well for one frame and partially good for the other why the cars in the left half of the frame were not detected at al. I tried to cut off the frame and apply the model on the cropped frame separately, but the performance didn't change at all. Do you have any idea?

zungam commented 6 years ago

I see this is the same image on all pictures, just cropped in different ways. I also notice that the cars which were not detected, is how dark they are. It could be that the model is not trained for dark images of cars. Try increasing the brightness in the image in these areas and see what happens.

Is this the coco trained model, or your own? If it is your own, how much data did you use? How different where the data? Did you use data augmentation?

baharna2 commented 6 years ago

Thanks for your reply. 1- it's coco model not my own. 2- The first two pictures are taken from two cameras installed on the roof of the building. The rest are just the cropped images from the first camera feed. 3- Is there any specific method to increase the brightness?

Thanks for your support, Bahareh

Paulito-7 commented 6 years ago

@baharna2

Look at the plugin imgaug which has been interfaced to Mask RCNN, it allows you to simply generate brighter images on the fly when training by adding an augmenter in model.py --> load_gt_image as follows : augmenter = iaa.Sequential([ iaa.Sometimes(0.5, iaa.Multiply((0.6, 1.4)) ), ],) Please have a look here for further information

padmaksha18 commented 5 years ago

@zungam HI Magnus, I am using this instance segmentation for an use case where we are masking personal identification details like face or any number plates etc. So for that , i want to make the masks more intense like it should be completely opaque on the face or car number plate. Can you please let me know where in the visualize.py i should change? The colors are generated with the random_colors function though.Many thanks in advance.