Active-Learning-For-Object-Detection

Active Learning For Object Detection

Computes informativeness of a single object detection using MC Dropout and DBSCAN. After computing the model’s uncertainty per detection, it aggregates the individual uncertainty values in two different ways. The first metric selects the images with crowded scenes, while the other is neutral to the number of detections in an image. Selection strategy: maximizing uncertainty. Dataset: KITTI

1. Create tfrecords: /object_detection/dataset_tools/ prepare_kitti_splits.py create_kitti_tfrecord_init_and_test.py create_kitti_tfrecord_random.py

2. Train & evaluate: /object_detection/legacy/train.py and /object_detection/legacy/eval.py

3. Config file: /object_detection/faster_rcnn_resnet101_kitti.config In this file 'use_dropout: true' and 'dropout_keep_probability: 0.5'

4. Export the trained graph: /object_detection/export_inference_graph.py. A flag --mc_dropout must be True. With this flag the graph will keep the dropout layers for inference

5. Compute uncertainties: /object_detection/compute_image_uncertainties.py.

6. Create tfrecords for the most informative images with /object_detection/dataset_tools/create_kitti_tfrecord_uncertainty.py