PoCo: polyp detector/counter prototype ====================================== This folder contains datasets and detector data for the prototype implementation of the PoCo polyp detector/counter [1,2]. Corresponding code and installation instructions: http://github.com/rokm/polyp-detector 1. References: -------------- [1] M. Vodopivec et al., Polyp counting made easy: two stage scyphistoma detection for a computer-assisted census in underwater imagery, Fifth International Jellyfish Bloom Symposium: Abstract book, Barcelona, 2016 [2] M. Vodopivec et al., Towards automated scyphistoma census in underwater imagery: useful research and monitoring tool, 2018 2. Contents (original version): ------------------------------- Two versions of the data is provided. The one used by the experimental code and original journal submission is described in this Section. During the paper revision, some of the datasets have been reorganized and split - those are described in Section 3. This folder contains the following data for replication of results in our journal submission using the experimental code: 2.1. detector-data.tar.xz ~~~~~~~~~~~~~~~~~~~~~~~~~ Default detector data, providing the pre-trained ACF detector and R-CNN's feature extractor network. See enclosed README for details. 2.2 dataset-kristjan.tar.xz ~~~~~~~~~~~~~~~~~~~~~~~~~~~ Original dataset used by K. Shirgoski in his Bachelor's thesis. Contains 35 images of polyps, with anotated ROI polygons and bounding box annotations that were made by Kristjan himself. In addition, several hand-cut patches containing no polyps (i.e., hard-negative samples) are provided in a separate folder. The following split was originally defined (this split is also used by default in the vicos.PolypDetector class, e.g. if built-in train_and_evaluate() method is called without any additional parameters): * training images: 07.03.jpg, 13.01.jpg, 13.03.jpg, 13.04.jpg, 13.05.jpg * test images: 01.01.jpg, 01.02.jpg, 01.03.jpg, 01.04.jpg, 01.05.jpg, 02.01.jpg, 02.02.jpg, 02.03.jpg, 02.04.jpg, 02.05.jpg, 03.01.jpg, 03.02.jpg, 03.03.jpg, 03.04.jpg, 03.05.jpg, 04.01.jpg, 04.02.jpg, 04.03.jpg, 04.04.jpg, 04.05.jpg, 05.01.jpg, 05.02.jpg, 05.03.jpg, 06.01.jpg, 06.02.jpg, 06.03.jpg, 07.01.jpg ,07.02.jpg, 08.01.jpg, 08.03.jpg In first part of journal experiments, we use the negative patches from this dataset when training the ACF detector. In the second part of journal experiments (the feasibility study), we use ACF detector that was trained on this dataset or train one from scratch using the training images and negative patches from this dataset. 2.3 dataset-martin.tar.xz ~~~~~~~~~~~~~~~~~~~~~~~~~ The main dataset used in our journal experiments. It provides seven base images (five of which are taken from Kristjan's dataset) on which the first part of journal experiments were performed, plus eleven additional images that served as training data for the second part of experiments (the feasbility study). For each image, ROI polygon and bounding box annotations are provided (*.poly and *.bbox files, respectively). The first seven images: 01.01.jpg 02.02.jpg 02.04.jpg 05.01.jpg 07.03.jpg 100315_TMD_007.jpg 100315_TMD_022.jpg also provide the manual annotations made by different human annotators (*.txt files). These seven images were used to evaluate the consistency of annotators (Table 1 and Table 2) in journal submission, and to perform leave-one-out evaluation of the automatic detector (Table 3 and Table S1). For the additional eleven images: 120315_TMD_007.jpg 120413_TMD_011.jpg 120413_TMD_022.jpg 120514_TMD_007.jpg 120706_TMD_019.jpg 120914_TJASAp_017.jpg 120914_TJASAp_021.jpg 121024_TMp_020.jpg 121113_TMD_003.jpg 121113_TMD_018.jpg 130118_TMD_008.jpg only polygons and bounding boxes are provided. All eighteen images are used in the second part of journal experiments to train the detector, which is then used to obtain detections on the third dataset (dataset-sara), from which the polyp density is estimated (Figures 5 and 6 in the journal submission). 2.4. dataset-sara.tar.xz ~~~~~~~~~~~~~~~~~~~~~~~~ The series of images from (Hocevar et al., 2016) that were used as test data for the second part of the journal experiments (the feasibility study). For each image, the ROI polygon and a text file with manual human annotations are provided. NOTE 1: the annotations are result of a single-pass manual annotation, and as such cannot be considered proper ground-truth. However, they may still be useful to compare the detections from the automatic detector to those made by the human annotator. NOTE 2: the images were captured without considering the automatic processing. Therefore, there are significant variations in image scale (both in image resolution and in viewpoint/zoom). Several images may have been taken under poor lighting conditions. The info.mat file contains bounding box annotations of few (10~50) representative polyps in each image. This information is used to estimate distance thresholds in automatic evaluation (i.e., to find assignments between detections and annotations), and may be also used to compensate for variations in image scale (resolution and zoom level). 3. Contents (revised version): ------------------------------ During the revision of paper submission, dataset-sara has been renamed to DatasetB, while dataset-martin has been split into DatasetA (first seven images) and DatasetB_ExtraSamples (the additional 11 training images). Thus the contents are as follows: 3.1. DatasetA.tar.xz: ~~~~~~~~~~~~~~~~~~~~~ "Dataset A" from paper [2]. This corresponds to the first seven images from dataset-martin.tar.xz, along with their annotation data. 3.2. DatasetB_ExtraSamples.tar.xz: ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Extra training samples from dataset-martin.tar.xz, which are used to train the detector for experiment on "Dataset B" in the paper [2]. 3.3. DatasetB.tar.xz: ~~~~~~~~~~~~~~~~~~~~~ The new name for dataset-sara.tar.xz. Contains the series of test images for the feasibility study part of the experiment in the paper [2]. 3.4 Figure1_originals.zip ~~~~~~~~~~~~~~~~~~~~~~~~~ Source images for Figure 1 in the paper [2].