FishFeeding-NB: A Benchmark Dataset and Spatiotemporal Models for Fish Feeding Activity Recognition in Marine Floating Cages

FishFeeding-NB: Marine Cage Aquaculture Fish Feeding Video Dataset & Benchmark Package FishFeeding-NB is a benchmark dataset and deep learning toolkit for binary fish feeding activity recognition in commercial marine floating cage environments (Nha Phu Bay, Khanh Hoa, Vietnam). Acquired continuously via an edge-deployed AIoT system using a cage-mounted top-view IP camera. Dataset Overview: Total Video Clips: 4,934 standardized 2.0-second video clips (1080p, 25 fps, H.264/AVC). Binary Classes: 0_normal (3,436 clips): Calm water, free swimming, intense solar glare, surface wave turbulence (hard negative samples). 1_feeding (1,498 clips): Active fish feeding frenzies, intense surface splashes, white foaming, and feed pellet consumption. Data Split: Session-disjoint and date-disjoint 3-way partition: Train set: 3,129 clips (2,164 normal, 965 feeding across 8 dates) Validation set: 633 clips (473 normal, 160 feeding across 2 dates) Test set: 1,172 clips (799 normal, 373 feeding across 3 dates), containing a dedicated 460-clip hard-negative test subset (challenging solar glare & wave turbulence). Synchronized Multimodal Telemetry: Accompanied by synchronized 57-session (9.57 hours) underwater hydrophone acoustic recordings (48 kHz, 16-bit) and multiparameter water-quality telemetry. Package Contents: videos/: Complete 4,934 standardized video clips organized by split and class. annotations/: Split CSV lists, clip labels, hard-negative subset definitions, and acoustic session indices. models_and_benchmark/: Complete PyTorch inference and benchmark evaluation pipeline, model definitions (MobileNetV3, ResNet-18, EfficientNet-B0, R3D-18, MC3-18, R(2+1)D-18), and best pretrained model weights (.pth). README.md: Comprehensive usage instructions and benchmark replication guide.