MediaEval 2013: Placing Task
I participated in the MediaEval 2013 along with a classmate and a professor as our guide. The task we chose was the Placing Task. This involved developing a system that can estimate the geographic location where a photograph was taken based on it's tags.
We did this by dividing the world up into a grid and assigning images to cells of this grid. Initially, the entire world is one cell and all images are put into this cell. Once the number of images in the cell exceeds 10,000, the cell is divided into 4 equal sized cells and the images are reallocated into the smaller cells. This goes on until all the images have been assigned to a cell.
We did this by dividing the world up into a grid and assigning images to cells of this grid. Initially, the entire world is one cell and all images are put into this cell. Once the number of images in the cell exceeds 10,000, the cell is divided into 4 equal sized cells and the images are reallocated into the smaller cells. This goes on until all the images have been assigned to a cell.
Now, we created a bag-of-words language model for each cell based on the most frequently occurring tags. Using this, we were able to assign the right cell to a new image using its tags pretty accurately.
Once the most probable cell is found, the image is given the coordinates of an image from the grid which has maximum number of common tags.
This worked fairly well - not the best results by far, but a good start nonetheless. The model was tested against 5300 images and accuracies were tested with different allowances for error. Here are the results:
Error allowance Accuracy
1km 0.74%
10km 3.9%
100km 15.24%
500km 26.3%
1000km 30.14%
The results were published in the proceedings of the MediaEval 2013 conference. Check out the paper here.
Once the most probable cell is found, the image is given the coordinates of an image from the grid which has maximum number of common tags.
This worked fairly well - not the best results by far, but a good start nonetheless. The model was tested against 5300 images and accuracies were tested with different allowances for error. Here are the results:
Error allowance Accuracy
1km 0.74%
10km 3.9%
100km 15.24%
500km 26.3%
1000km 30.14%
The results were published in the proceedings of the MediaEval 2013 conference. Check out the paper here.