How to read the output of the graph?

[noob-ja]

hi, so how exactly we read the output of the network?

for example, using yolov3-tiny, i feed in an image the output of it is with shape (1,26,26,255)

whats the next step to get the predicted class and bounding box? sry im new with this

[noob-ja]

im able to extract the output now, however, the result is not what i was expected

it keeps detecting a lot of aeroplane and apples in every single photo i feed in

in the picture below, i have set the object threshold to be 0.7, yet it detected aeroplane and apples to be above even that

[sjain-stanford]

@noob-ja - the purpose of DW2TF is to translate the convolutional layers in a network defined in darknet to TensorFlow. It is up to you to add the necessary post-processing (e.g. non-max suppression etc) to convert the features to the respective bounding box and class predictions. Sorry I couldn't be of help here.

[jinyu121]

@noob-ja After getting the output of the net, use non-max suppression (NMS) to select the "best" boxes. The pipeline in original YOLO code is "get the output of the net -> set a threshold to filter the boxes -> NMS". For the last image, may be the net is not well-trained? Or, you should adjust the clustering parameters before training.

[noob-ja]

i see, thx for the help :) really appreciate it

[kapil0kumar]

@noob-ja have you found the way to perform nms on the output?

[nishanthcgit]

@noob-ja sorry, I am quite new to TF, how did you get the bounding box coordinates from those output layers? I am aware of how to do nms etc but can't understand how you can convert the output layers into bboxes,classes and scores (apologies if this is a basic question)

[Robin2091]

@kapil0kumar Hi, were you able to extract the bounding boxes from the output? I'm having some trouble with that.