See-with-Sound

Code:

Below are the step to setup the code and perform training

Setup:

After setting up the code as below, update the paths appropriately

git clone https://github.com/ksasi/See-with-Sound.git

Install Dependencies:

cd See-with-Sound

pip install -r requirements.txt

Dataset:

• Download Food-101 dataset
• Execute ipython notebook Minor_Project_Data_Curation.ipynb to generate audio samples for each category
• Execute ipython notebook Minor_Project_DataSet_Setup.ipynb to setup food-101-small dataset
• The setup food-101-small dataset consists of Train, Probe, Gallery and Other folders

Dataset setup food-101-small structure :

food-101-small/
              Train/
                  <class_name>/
                            <image_id1>.jpg
                            <image_id2>.jpg
                            ...
                            ...
                            ...
              Probe/
                  <class_name>/
                            <class_name>.wav
              Gallery/
                  <class_name>/
                            <image_id1>.jpg
                            <image_id2>.jpg
                            ...
                            ...
                            ...
              Other/
                  No_Image_Available.jpg

Training:

After updating the paths, train SGDClassifier incrementally as below :

nohup python model_train.py &

Evaluation:

The trained model can be evaluated as below :

nohup python evaluate.py &

Results:

The CMC Curve of Probe and Gallery is shown below :

Rank1 Identification Accuracy: 87.097%

Demo:

Demo of Image search from audio input can be executed by running Audio_Image_Search_Demo.ipynb ipython notebook

References

The code is adapted from the following repositories:

• Low Resolution face recognition - Github Link