tensorflow Object detection, Bounding Boxes are not visible on the own dataset

Question

I am trying the run the TF object detection locally on my own dataset. Every step is happening perfectly except the Visualization of the BB on a test image. First I have run the Pascal VOC dataset on the Faster R-CNN Inception ResNet v2, modified the scripts as per the VOC dataset and Then followed the instructions from G3doc, everything worked perfectly, for the visualization, I am using the ipython jupyter notebook given in the objection detection. The visualization was awesome. Then I tried to do the same for my own dataset. Repeated all the steps same but no BB is showing on the image.

Can someone help what might be going wrong?

PS I am using ubuntu 16.04, 64GB ram system.

Answer 1

Try passing the bounding boxes to a print statement. Do you get any valid output? If so then it is probably nothing to do with your model and your bounding boxes are created just fine. What is your mAP? If it is very early on in the training process then your accuracy is probably too poor and do not meet a minimum threshold.

Take a look at the visualize_boxes_and_labels_on_image_array() function from object_detection/utils/visualization_utils.py, note the default value for min_score_thresh . You can either change this default value or pass in min_score_thres=0 as an argument when you call the function.