search

cvlab-epfl / detecting-the-unexpected

Detecting the Unexpected via Image Resynthesis
Other
56 stars 10 forks source link

In the detection of unknown items, the detection results are shown in the figure below.Is the bucket not detected because the data set is different? #13

Open bodhi-tree01 opened 3 years ago

bodhi-tree01 commented 3 years ago

In the detection of unknown items, the detection results are shown in the figure below. Is the bucket not detected because the dataset is different? If I use my data set for training, can I detect unknown objects?Roughly how many images are needed for the dataset? Looking forward to your reply. Thank you very much！ ZU3KHLOM`RZKZZXD1 VB3XI

adynathos commented 3 years ago

Hello,

When the image and detection heatmap is overlaid, the bucket appears to be detected. Also the walls of the tunnel get marked because they are unusual with respect to the Cityscapes training set.

If I use my data set for training, can I detect unknown objects?Roughly how many images are needed for the dataset?

This network was trained with Cityscapes (3k images) - training anomalies were synthetically created by altering the provided labels of selected objects. This kind of training requires the dataset to be have per-pixel semantic labels.

If you want to detect items specifically on the road, the training images can be created by synthetically injecting objects on the road surface, which is described in our later work.

resynthesis_github_merge

bodhi-tree01 commented 3 years ago

Thank you very much for your reply and wish you all the best in your work