This repo contains the work in progress Deep Learning part of NatCap's Detecting Dams with AI and Satellite Imagery project. It is funded by the National Geographic and Microsoft AI for Earth program.

Imagery

1. Labeling data: merged xx databases, and bounded xx dams (with http://natgeoserver.ecoshard.org:8888/).
2. Fetching Planet Imagery: https://bitbucket.org/natcap/extract-dams/src/default/
3. Offset imagery so dam isn't always centered: (Maanas) https://github.com/charlottegiseleweil/dams/blob/master/utils/image_preprocessing/create_cropped_offset_imagery.ipynb

Training

1. Format Inputs (Make TFRecords) (see TFOD Preprocessing)
2. Train: see readme in dams/tfod/training/

python model_train.py --logtostderr --train_dir=training/ --pipeline_config_path=training/{{CONFIG FILE}}.config

Inference

1. Export frozen graph with dams/tfod/detection/export_inference_graph.py

python export_inference_graph.py \
    --input_type image_tensor \
    --pipeline_config_path ../../../../repos/dams/tfod/training/configs/07_29_imagery7-25_Faster_rcnn_resnet50_coco.config \
    --trained_checkpoint_prefix ../../../../outputs/fasterRCNN_07_27_newimagery/model.ckpt-300000 \
    --output_directory ../../../../outputs/fasterRCNN_07_27_newimagery/export_inference \
    --write_inference_graph True

2. Run detector inference.py

python inference.py
          --modelFile {path_to_frozen_graph.pb} 
          --imgDir {path_to_dir}
          --outputDir {path_to_dir}
          --bbox_format (Optional) {outputs bboxes format. Default is y1x1y2x2_pixel}
          --confidenceThreshold (Optional) writes only bboxes above this threshold. Default to 0.05
          --max_boxes_per_images (Optional) {int} Will write this # of bboxes per image. Default to 1

Playground/Visualize inference on a few images: Inference notebook

Evaluation

See Evaluation notebook

1. Store results table (a row per image), with Write_results_csv.py

2. Calculate custom Evaluation metric:

• Recall @ IoU 5% and Confidence > 0.1.
• for Small Dams (bbox 500-1800 pixels, i.e dam of ~meters)

3. Visualize images (to analyze mistakes etc..)

4. Human baseline performance

/!\ To fix:

1. Results table only retains 1 bbox/images, allow for several bboxes/img (must re-run Inference.py to get several bboxes/img)
2. Fix Custom eval metric (Category column)

---

# Samples

---

Other experiments

Single shot detectors (YOLOv3 with Darknet)

See dams/experimental/yolov3/readme_ultralytics.md