SHREC 2020: Track on river gravel characterization
General description
River beds are characterized by the composition of sands, gravels, cobbles in various shapes and sizes. Their distribution is quite important as it is linked to fluvial dynamics and related processes (e.g., hydraulic resistance, sediment transport and erosion, habitat suitability [1]). Ondulations of the bed surface, or bed forms, though having length scale usually larger than the sediment sizes, interplay with the grains and must be recognized.
Hydrologists are strongly interested in finding automatic and fast measurement methods for grain size distribution, as classical methods employed are time-consuming and labor-intensive.
Recently, image processing methods have been proposed for the automatic quantification of the grain size distrubution [2]. However, as 3D sensors or photogrammetric reconstruction can be used as well in underwater applications, it is interesting to understand if current shape analysis methods could possibly provide a reliable estimation of the gravel size distribution from a noisy acquisition of the bed shape.
For this reason we created a dataset including real captures of riverbed mockups and propose a related classification/retrieval task.
[1] Alessio Cislaghi, Enrico Antonio Chiaradia, Gian Battista Bischetti,
A comparison between different methods for determining grain distribution in coarse channel beds,
International Journal of Sediment Research,
Volume 31, Issue 2, 2016,
[2] Chung, Chang-Han, and Fi-John Chang. "A refined automated grain sizing method for estimating river-bed grain size distribution of digital images." Journal of Hydrology 486 (2013): 224-233.
Dataset creation
We created realistic mockups of river beds composed of gravels with known size ranges. 8 different mixtures of gravels with size selected in known ranges have been used:
- Class 1: 3-5 cm
- Class 2: 3-5 cm and 4-6 cm mixed
- Class 3: 3-5 cm and 6-10 cm mixed
- Class 4: 4-6 cm
- Class 5: 4-6 cm and 10-20 cm mixed
- Class 6: 6-10 cm
- Class 7: 6-10 and 10-20 cm mixed
- Class 8: 10-20 cm
These mockups have been captured with a digital camera from different viewpoint also capturing the box of known size. Photogrammetry has been used to create 3D models of the mockups. Untextured 3D models have been divided in patches, each one associated to the corresponding grain size composition. The resulting dataset is a set of 256 patches, representing 8 different classes of grain size distributions.
Images of bed mockups with different gravel size composition
Renderings of sample surface patches derived from photrogrammetric reconstruction of simulated riverbeds with different gravel size composition
Dataset Download
The dataset is now available (click here to download) . The archive includes two folders: Train, with 16 models belonging to known classes, and test, with 240 models with undisclosed classes and named "model_X.ply" with X from 1 to 240. Files are stored in binary ply format.
Task and evaluation
We designed two tasks:
- Classification: given the class features and two surface patches with known class, assign the correct class to the rest of the samples
- Retrieval: given the a full set of patches with unknown label, provide a dissimilarity matrix that can be use to retrieve most similar gravel patterns to a query one
Classification accuracy will be evaluated with accuracy and confusion matrices.
The evaluation of the retrieval algorithms will be performed using the precision-recall curves, nearest neighbor, first tier, second tier, normalized discounted cumulated gain and average dynamic recall. Participants can send up to 3 classification results and/or dissimilarity matrices.
Classification files will be plain text files with N=240 lines, and just the label(1-8) of the gravel class attributed to the corresponding test patch. An example file is linked here
Dissimilarity matrices will be plain text files with, for each line i the N sequential values of the dissimilarity of the patch i from the pathches numbered from 1 to N, separated by spaces. An example file is linked here
Timeline
- March 2: dataset available
- March 10: Registration deadline: participants should send an e-mail to register for the contest to andrea.giachetti(at)univr.it
- April 12: results sent by participants and collected together with methods descriptions
( one-page latex file with at most two figures )
- May 3: submission of the track report to the reviewers
Contacts
Andrea Giachetti (email: andrea.giachetti(at)univr.it ) - VIPS Lab, University of Verona
Filippo Andrea Fanni (email: filippoandrea.fanni(at)univr.it ) - VIPS Lab, University of Verona
Silvia Biasotti (silvia(at)ge.imati.cnr.it) - CNR IMATI, Genova
Elia Moscoso Thompson (elia.moscoso(at)ge.imati.cnr.it) - CNR IMATI, Genova
Luigi Fraccarollo (luigi.fraccarollo(at)unitn.it ) - DICAM University of Trento
