Video Data - Fish Species Classification

Icelandic River Fish Counter Video Dataset Abstract This dataset contains 9,257 underwater video recordings from automated fish counters deployed in three Icelandic rivers (Krossá, Laxá in Leirársveit, and Selá) between 2018-2022. Videos were captured using Vaki Riverwatcher systems with motion-triggered recording and contain three native salmonid species: Atlantic Salmon (Salmo salar), Brown/Sea Trout (Salmo trutta), and Arctic Char (Salvelinus alpinus). Each video is expert-labeled and represents controlled underwater monitoring conditions suitable for computer vision and automated species classification research. This dataset was put together as part of the study [X]. This includes all the video data used for the ful analysis and modelling as well as the data to do a quick analysis, ideal for reproducing the paper. To better understand how make the most out of this dataset we reccomend consulting the respective paper and Github page. Video Dataset Composition Total Videos: 9,257 Species Distribution: Trout: 3,252 (35.1%) Salmon: 2,233 (24.1%) Char: 866 (9.4%) Not fish (false positives): 2,906 (31.4%) Collection Sites: Krossá, Laxá in Leirársveit, Selá rivers Time Period: 2018-2022 Reproducibility Ready Dataset A dataset that is ideal for someone that is interested in reproducing the study results while jumping over the heavy computational and time consuming parts (getting detection scores and extracting features). Detection Scores: Detection-Scores-ViT-SO400M-14-SigLIP.pkl Scores-ViT-SO400M-14-SigLIP-MultipleFish.pkl Feature Extraction: ResNet-50-features ViT-SO400M-SigLIP-features video_df_complete.csv: mapping between species in icelandic and videos paths. Detection Experiments Data Data used to validate and test the detection datasets validation_set: fish - frames containing fish no_fish - frames not containing fish test_set_final fish no_fish validation_fish_counting: more_than_one_fish - frames with multiple fish one_fish - frames with a single fish Technical Specifications Video Format: MP4 video files Resolution: Primarily 640×400, some 640×360 and 640×480 Frame Rate: Mostly 25 fps (with exceptions, e.g., few videos with 9.8 fps and 7.7 fps) Recording: Color during daylight, infrared B&W at night Trigger: Motion-activated via infrared scanner Data Quality Videos include diverse conditions: clear water, turbid water, low light, biofouling, debris, and both single and multiple fish scenarios. The high false positive detections (~31% of non-fish videos) reflects real-world monitoring challenges. All species classifications performed by expert ichthyologists at the Marine and Freshwater Research Institute of Iceland. Applications Suitable for fish species classification, underwater computer vision, automated monitoring systems, ecological research, and algorithm development for aquatic environments. Data Provider Marine and Freshwater Research Institute (MFRI), Iceland Usage Notes Each video has a single species label corresponding to the first fish that triggered the motion sensor. Videos may contain environmental artifacts and varying visibility conditions typical of real-world aquatic monitoring systems.