Deep-Learning-for-Dendritic-Spines-Detection

Benchmarking Yolov2, Faster-RCNN and Shape-Priors-CNN on dendritic spines detection

Please read the report for more details

1-Faster-RCNN:

The model is already trained.

• The final weights are in Faster-RCNN/inference_graph.

• The results are in Faster-RCNN/results

• To re-calculate the predictions on the test set run F1_score_and_predictions.py

• To recalculate on other images change the path to the folder containing the images.

2-SP-CNN:

run the jupyter notebook. Note that the train images are not complete (due to size issues), but they will give you a sense of the overall pipeline.

3-YOLOv2:

Testing:

• create a text file and name it test.txt. this file contain the path to images you want to test on. Get inspired by an already existing test.txt and train.txt to get an idea.

• To calculate the map, f1-score... at a specific threshold (say 0.5, change it to other values for precision-recall graph) do the following:

From darknet folder run :./darknet detector map cfg/obj.data cfg/yolo-obj.cfg backup/yolo-obj_last.weights -thresh 0.5

• To test on a specific image (say the image's name is img.png) do the following:

From darknet folder run: ./darknet detector test cfg/obj.data cfg/yolo-obj.cfg backup/yolo-obj_last.weights img.png -thresh 0.55

References:

• You Only Look Once: Unified, Real-Time Object Detection
• YOLO9000: Better, Faster, Stronger
• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
• Deep Networks with Shape Priors for Nucleus Detection