Oil palm ( Elaeis guineensis ) plantation yield is estimated periodically through the Black Bunch Census (BBC), a field practice in which trained staff count and classify fruit bunches to project output for the coming months. Photographing each tree from multiple positions captures all visible bunches but introduces a cross-view deduplication problem: the same bunch appears in several images, and naive summation of detections far exceeds the true unique count. No publicly available oil palm dataset addresses this problem with annotated ground truth. SawitMVC contains 3,992 smartphone images of 953 oil palm trees collected from two commercial plantations in Kabupaten Tanah Laut, Kalimantan, Indonesia (DAMIMAS plantation: 854 trees; LONSUM plantation: 99 trees), each photographed from four sides at 90-degree intervals or eight sides at 45-degree intervals. Bunches are annotated across four BBC maturity classes by estimated months until harvest: B1 (≤1 month), B2 (∼2 months), B3 (∼3 months), and B4 (∼4 months), in YOLO bounding box format for detection and classification tasks. A separate per-tree JSON ground truth layer records 9,823 unique bunches with their cross-view appearance links, supporting direct evaluation of counting algorithms independently of detection performance.
Indriani et al. (Mon,) studied this question.