How to filter only stop sign and pedestrian

[sudonto]

Great work! Btw, I want to train the model but stuck in this step:

"Follow instructions in the LISA Traffic Sign Dataset to create 'mergedAnnotations.csv' such that only stop signs and pedestrian crossing signs are shown"

How do I filter all except stop sign and pedestrian cross? I read the mergeAnnotationFiles.py but no clue. Please help me.

[YashBansod]

Can someone please specify the arguments to pass to the mergeAnnotationFiles.py for doing this?

While the pwd being the $LISA/tools directory of the dataset I tried:

• python mergeAnnotationFiles.py frame mergeAnnotations.csv ../
• python mergeAnnotationFiles.py frame mergeAnnotations.csv ../aiua120214-0/frameAnnotations-DataLog02142012_external_camera.avi_annotations
• python mergeAnnotationFiles.py video mergeAnnotations.csv ../aiua120214-0/frameAnnotations-DataLog02142012_external_camera.avi_annotations

but the output always comes: No annotation files found, exiting...

please help me out.

[YashBansod]

Okay I found a way of at least running the mergeAnnotationFiles.py.

I copied it to the $LISA folder and ran the following code:
python mergeAnnotationFiles.py frame mergeAnnotations.csv

this merged all the annotations from frames present at any sub directory of the $LISA directory. But I still cannot find a way to merge only annotations of stop signs and pedestrian crossing signs.

[sudonto]

@YashBansod,

I think we have to modify mergeAnnotation.py such that only signs we want to put in csv file.

[YashBansod]

@sudonto if you were able to filter only stop and pedestrian signs please tell how you did it.

I am currently able to extract only pedestrian and stop signs using:
python extractAnnotations.py -f stopAhead -f pedestrianCrossing copy ../allAnnotations.csv

this creates a folder named annotations which will contain all the pics where pedestrianCrossing and stopAhead signs are available.

But I am still unable to build the annotations for them

I even tried:
python mergeAnnotationFiles.py frame mergedAnnotations.csv annotations/
still I find that the mergedAnnotations.csv is the same as allAnnotations.csv

[sudonto]

@YashBansod,

I will let you know soon I have the solution.

[sudonto]

Hey @YashBansod ,

I was able to build only stop sign and pedestriancrosssing sign into one csv file. In my case, I use filterAnnotationFile.py first to generate those signs into seperate folder and then execute mergeAnnotationFiles.py to combine 2 csv file (stop and pedestrian sign) into one file.

Please try.

[YashBansod]

Hi @sudonto
Sorry for the late reply. I was able to get this thing working. In issue #21 steps mentioned by @jacksmith21 are sufficient to run get the code working. the mergeAnnotationFiles.py just merges all the annotations in all the sub-directories, but the dataprep.py handles the seperation of the pedestrian and stop sign images. The documentation of this was a bit misleading, causing a lot of my time to be wasted just on figuring on how to make this work. Anyway the output was not satisfactory for my application and I couldn't replicate the outputs mentioned in the readme of this repository, so I started looking elsewhere.

[sudonto]

@YashBansod ,

Yes, you absolutely correct about this. I started to analyze each code of dataprep.py week ago and I found that there is a line of code that filters out the annotation tag other than the desired signs.

Now, I have 2 problems afterward: my loss function is very large (arround 200 after 200 epochs) and the produced model always give me an error in inference mode (division by zero when calculating the IoU, sighhh).

I think I'd like to search for another SSD algorithms.

PS.: In case you find another robust SSD algorithm yet offer simple code, would you mind to share with me? :)

[YashBansod]

@sudonto I didn't get any div by zero errors and don't really understand why its happening for you, but you can try re cloning the repo and do the changes I mentioned in #27 and it should work. Also, my loss was even bigger. It was around 500 after 200 epochs and around 400 after 2000 epochs. Furthermore, in tensorboard, I could see that my model was over fitting. Instead of working on SSDs, which I was even testing in the first place as an alternative to Faster RCNNs, I would rather put my effort on YOLO v2. Anyway, am currently involved in a different project but if I do find a good SSD implementation, I will post it here.

[YashBansod]

@sudonto were you able to run the python inference.py -m demo with the pre-trained model provided by the creator of this repo?

[sudonto]

Yes, I was able to run python inference.py -m demo with success using pre-trained model, showing images in sample_images folder with bounding box (no division by zero).