Wheatears counting in images
Counting wheatears from images using deep learning
You can find more details in the paper and the colab notebook below.
- The paper
- Colab URL
We are supplied with a raw set of images of a wheatfield with a flock of wheatears. We want to count the number of wheatears in each image. This is an actual problem that requires a lot of manual work from farmers. Apparently, wheatears count is a great indicator of the health of the wheat field.
To accomplish this, we used two algorithms:
While EfficientDet is a better model than Faster R-CNN, we wanted to compare the two models to see how they perform on this task.
Dataset: We used the Global Wheat Dataset for training and testing the models.
Some results
$$ \begin{array}{ |c|c|c|c|c|c| } \hline ImageID & GT & Detected & Precision & Recall & Accuracy \\ \hline 2fd875eaa & 27 & 24 & 1.0 & 0.89 & 88.9\% \\ 51b3e36ab & 27 & 29 & 0.86 & 0.93 & 80.6\% \\ 51f1be19e & 18 & 18 & 1.0 & 1.0 & 100.0\% \\ 53f253011 & 31 & 29 & 1.0 & 0.94 & 93.5\% \\ 348a992bb & 37 & 36 & 0.97 & 0.95 & 92.1\% \\ 796707dd7 & 31 & 23 & 1.0 & 0.74 & 74.2\% \\ aac893a91 & 24 & 21 & 0.95 & 0.83 & 80.0\% \\ cb8d261a3 & 24 & 21 & 1.0 & 0.88 & 87.5\% \\ cc3532ff6 & 26 & 29 & 0.9 & 1.0 & 89.7\% \\ f5a1f0358 & 28 & 31 & 0.9 & 1.0 & 90.3\% \\ \hline Total & 273 & 261 & 0.95 & 0.91 & 87.4\% \\ \hline \end{array} $$
Faster R-CNN with SGD Optimizer on Test Data
$$ \begin{array}{ |c|c|c|c|c|c| } \hline ImageID & GT & Detected & Precision & Recall & Accuracy \\ \hline 2fd875eaa & 27 & 24 & 1.0 & 0.89 & 88.9\% \\ 51b3e36ab & 27 & 29 & 0.9 & 0.96 & 86.7\% \\ 51f1be19e & 18 & 18 & 1.0 & 1.0 & 100.0\% \\ 53f253011 & 31 & 29 & 1.0 & 0.94 & 93.5\% \\ 348a992bb & 37 & 36 & 0.97 & 0.95 & 92.1\% \\ 796707dd7 & 31 & 25 & 1.0 & 0.81 & 80.6\% \\ aac893a91 & 24 & 21 & 0.95 & 0.83 & 80.0\% \\ cb8d261a3 & 24 & 21 & 1.0 & 0.88 & 87.5\% \\ cc3532ff6 & 26 & 29 & 0.9 & 1.0 & 89.7\% \\ f5a1f0358 & 28 & 31 & 0.9 & 1.0 & 90.3\% \\ \hline Total & 273 & 263 & 0.96 & 0.92 & 88.7\% \\ \hline \end{array} $$
Faster R-CNN with Adam optimizer on Test Data
$$ \begin{array}{ |c|c|c|c|c|c| } \hline ImageID & GT & Detected & Precision & Recall & Accuracy \\ \hline 2fd875eaa & 27 & 24 & 0.88 & 0.88 & 88\% \\ 53f253011 & 31 & 30 & 0.96 & 0.96 & 96\% \\ 51b3e36ab & 27 & 25 & 0.92 & 0.92 & 92\% \\ 51f1be19e & 18 & 18 & 1.0 & 1.0 & 100\% \\ 348a992bb & 37 & 35 & 0.94 & 0.94 & 94\% \\ 796707dd7 & 31 & 26 & 0.83 & 0.83 & 83\% \\ aac893a91 & 24 & 21 & 0.87 & 0.87 & 87\% \\ cb8d261a3 & 24 & 24 & 1.0 & 1.0 & 100\% \\ cc3532ff6 & 26 & 25 & 0.96 & 0.96 & 96\% \\ f5a1f0358 & 28 & 28 & 1.0 & 1.0 & 100\% \\ \hline Total & 273 & 257 & 0.92 & 0.93 & 92.7\% \\ \hline \end{array} $$
EfficientDet-D5 results on Test Data
$$ \begin{array}{ |c|c|c|c|c|c| } \hline ImageID & GT & Detected & Precision & Recall & Accuracy \\ \hline 2fd875eaa & 27 & 24 & 0.88 & 0.88 & 88\% \\ 53f253011 & 31 & 30 & 0.96 & 0.96 & 96\% \\ 51b3e36ab & 27 & 25 & 0.92 & 0.92 & 92\% \\ 51f1be19e & 18 & 18 & 1.0 & 1.0 & 100\% \\ 348a992bb & 37 & 38 & 0.97 & 1.0 & 97\% \\ 796707dd7 & 31 & 26 & 0.83 & 0.83 & 83\% \\ aac893a91 & 24 & 19 & 0.79 & 0.79 & 79\% \\ cb8d261a3 & 24 & 24 & 1.0 & 1.0 & 100\% \\ cc3532ff6 & 26 & 25 & 0.92 & 0.96 & 92\% \\ f5a1f0358 & 28 & 28 & 1.0 & 1.0 & 100\% \\ \hline Total & 273 & 257 & 0.92 & 0.93 & 92.7\% \\ \hline \end{array} $$