News: the training datasets have been released on March, 13th 2017
The SeaCLEF 2017 task originates from the previous editions (2014 and 2015,2016) of marine organism identification in visual data for ecological surveillance and biodiversity monitoring. SeaCLEF 2017 significantly extends past editions in the tackled marine organisms species as well in the application tasks.
The need of automated methods for sea-related multimedia data and to extend the originary tasks is driven by the recent sprout of marine and ocean observation approaches (mainly imaging - including thermal - systems) and their employment for marine ecosystem analysis and biodiversity monitoring. Indeed in recent years we have assisted an exponential growth of sea-related multimedia data in the forms of images/videos/sounds, for disparate reasoning ranging from fish biodiversity monitoring to marine resource managements to fishery to educational purposes. However, the analysis of such data is particularly expensive for human operators, thus limiting greatly the impact that the technology may have in understanding and sustainably exploiting the sea/ocean.
Data collection: the SeaCLEF 2017 multimedia dataset contains images and videos of marine organisms. In all cases, the goal is to identify species or individuals from either images or videos.
In terms of tackled species, SeaCLEF will involves about 150 marine animals species from all over the world.
In particular the SeaCLEF datasets contains:
1) A set of videos recorded in Taiwan coral reefs for automated fish detection and species recognition;
2) A set of videos recorded near a water turbine for automated detection of rare events involving the presence of salmons;
3) A set of images of 148 marine animal species, weakly labelled.
4) A set of image pairs of of whale caudals.
Tasks description and datasets
According to the data available in the SeaCLEF 2017 contest, we defined four tasks:
Task 1: Automated Fish Identification and Species Recognition on Coral Reef Videos
The participants have access to the training set consisting of twenty underwater videos. The goal of the task is to automatically identify and recognize fish species. The fish species present in all videos as well as bounding boxes and fish species labels are
also provided in the training data.
Task 2: Automated Frame-level Salmon Identification in Videos for Monitoring Water Turbine
For this case, the participants have access initially to the training set consisting of eight underwater videos. The goal of the task is to identify in which frames salmon appear. Such events are pretty rate and salmons are often very small, thus the task mainly pertains detection of rare events involving unclear objects (salmons). Frame-level annotations indicating the presence of salmons are given along with salmon videos.
Task 3: Marine Animal Species Recognition using Weakly-Labelled Images and Relevance Ranking
Contrary to the previous tasks, this one aims at classifying marine animals from 2D images. The main difficulties of the task are: 1) high similarity between species and 2) weak annotations, for training, gathered automatically from the Web and filtered by non-experts. In particular, the training dataset consists of up to 100 images for each considered species (in total 148 fish species). Training images are weakly labelled, i.e., web images have retrieved automatically from the web using marine animal scientific names as query. The retrieved images were then filtered by non-experts who were instructed to only remove images not showing fish/marine animals. Furthermore, the filename for each image provides information about the relevance (relevance ranking) to the query and must be used during training. For example, for a given species X, the retrieved images are named from 1 to 100, with 1 being the most relevant image and 100 the less relevant one.
Task 4: Whale Individual Recognition
This task will aim at automatically matching image pairs, over a large set images, of same individual whales (such as the 3 matches of Figure 1) through the analysis of their caudal fins. Indeed, caudal fin that is most discriminant pattern for distinguishing an individual whale from another. Finding the images that correspond to the same individual whale is a crucial step for further biological analysis (e.g. for monitoring population displacement) and it is currently done manually by human operators (hence a painful, error prone and definitely unscalable process).
Task participation and run format
Participants are not expected to provide runs for all tasks, but can choose one (or a combination) of them. In particular
1) Teams submitting runs for task 1 have to submit runs for task 2 as well and vice versa.
2) Task 3 and task 4 are independent.
The participants must provide a run file named as TeamName_TaskX_runY.XML where X is the identifier of the task and Y the identifier of the run (up to three runs per participant per task). Run formats for each task will be released soon.
Figure 1: illustration of 3 individual whale matches (each line correspond to two images of the same individual)