beniz/deepdetect

Deep Learning API and Server

dgtlmoon

@dgtlmoon

maybe it's not taking into consideration if the object is zoomed in/out, small/big relative to the image... I saw something in the deepdetect API docs for SSD scale object hmmmm?

aAhhuh!

ssd_neg_overlap    float    yes    N/A    max overlap of negative samples with positive samples (bbox), between 0 and 1, e.g. 0.5

right !

Emmanuel Benazera

@beniz

don't touch that

what is wrong, the predictions ?

dgtlmoon

@dgtlmoon

yes

Emmanuel Benazera

@beniz

do you have examples ?

dgtlmoon

@dgtlmoon

yes, see above

Emmanuel Benazera

@beniz

screenshot is too small, I dont see anything

dgtlmoon

@dgtlmoon

ahh ok, sorry, two sec

dark area = original, colour/light = predicted

Emmanuel Benazera

@beniz

I'm not sure I understand exactly, I guess you need to give more thorough detail, i.e. is that tshirt not detected, and how many of these in % of your test set + your API prediction call + info on your model.

dgtlmoon

@dgtlmoon

Emmanuel Benazera

@beniz

At minima, if map is 1.0, then this means there's a box that intersects the ground truth with more than 50% area, so if the images above are from your test set, you are doing something wrong. And you can start by playing with confidence_threshold value at predict time.

dgtlmoon

@dgtlmoon

confidence_threshold is already 0.95

that image is not in the test set

Emmanuel Benazera

@beniz

lower it to 0.3

dgtlmoon

@dgtlmoon

yeah maybe the highest 'confidence' result isnt necessarily the right one?, checking now

Prediction service request

{"service":"tshirt_info_detection","parameters":{"input":{"width":300,"height":300},"output":{"confidence_threshold":0.3,"bbox":true},"mllib":{"gpu":false}},"data":"base64...."}

{"status":{"code":200,"msg":"OK"},"head":{"method":"/predict","service":"tshirt_info_detection","time":1666.0},"body":{"predictions":[{"classes":[{"bbox":{"ymax":461.73895263671877,"xmax":786.7083129882813,"xmin":111.97721099853516,"ymin":0.0},"cat":"1","prob":0.9653725028038025},{"last":true,"bbox":{"ymax":303.8266296386719,"xmax":585.5281372070313,"xmin":229.73147583007813,"ymin":9.2813138961792},"cat":"1","prob":0.3033854067325592}],"uri":"0"}]}}

Emmanuel Benazera

@beniz

then you can pick the first one

dgtlmoon

@dgtlmoon

Yes, the first one is what i'm already using

that's the BBOX that's far too large

Emmanuel Benazera

@beniz

ah ok

try the second one out of curiosity

and what model is this

dgtlmoon

@dgtlmoon

"xmax":786,"xmin":111 butI would expect like xmin=214, xmax=600

ssd_300 is the model i'm using

Emmanuel Benazera

@beniz

mmm, try refinedet, maybe one day if your dataset is public, just put it somewhere.

dgtlmoon

@dgtlmoon

second one = the xmax/xmin is perfect, but the ymax is totally off (ymin is OK)

Yeah its sort of public data from the website anyway

Check out this debug mode, You can see the prob bboxes

so... unsure...

yeah... ok, trying refinedet_512, will paste in the predictions, perhaps the images have too much geography going on for ssd

Emmanuel Benazera

@beniz

or you are overfitted to the bone

dgtlmoon

@dgtlmoon

I think that's more likely.... not sure how to solve that tho

I checked all train+test images with MD5 to verify i'm not accidently reusing a train image and test - all fine, I tried lowering the learning rate.. hmm added a LOT more test data too

I think the test is too similar to the train, even tho they are not exactly the same == overfitting

         "mllib":{
           "template":"ssd_300",
           "nclasses": 2,
           "finetuning":true,
          "rotate": true,
          "mirror": true,
          "noise": {
            "all_effects": true,
            "prob": 0.01
          },
          "distort": {
            "all_effects": true,
            "prob": 0.01
          },
          "gpu": true
        }

rotate, mirror, some noise and distort all enabled..

not sure what else to try

I try prob with 0.35 on training effects :) :) :)

dgtlmoon

@dgtlmoon

no difference.. "map": 1.000000011175871, at iteration 500

dgtlmoon

@dgtlmoon

....time to verify my setup, using the same training calls, but with a different dataset.....

@beniz what about using dropout to remove some features in the network?

i'll try

dgtlmoon

@dgtlmoon

i can put my data on kaggle if its useful

i think my data's bbox is usually within about 80% of the actual image... maybe thats the problem

Where communities thrive

People

Repo info

Activity