brakuta/Datepalm-Instance-Segmentation

Code for date-palm crown instance segmentation with CNN, transformer and state-space (Mamba) backbones # Date-palm instance segmentation across sensors and scales Mask R-CNN with eleven interchangeable backbones — CNN, transformer and state-space (Mamba-family) — trained and evaluated on date-palm crown delineation in imagery from **5 cm UAV down to 30 cm satellite**, and the adaptation that turned the best of them into a country-scale inventory. A fork of MMDetection 3.3.0 with project configs, backbone wrappers and tooling added. --- ## Start here: which experiment do you want? The work is five experiments, each answering a different question. **Pick the row you care about** — you do not need the others. | | question | imagery | folder | |---|---|---|---| | **1** | Which backbone is best when the sensor is fixed? | UAV 5 cm | `1_single_sensor_uav_5cm/` | | **2** | Does pooling two 15 cm sources help? | GE + aerial, 15 cm | `2_pooled_15cm_ge_aerial/` | | **3** | One model for all sensors, or one per sensor? | UAV 5 cm + GE + aerial | `3_unified_multisource/` | | **4** | How far does it carry to 30 cm, and at what labelling cost? | WorldView-3 30 cm | `4_satellite_wv3_30cm/` | | **5** | Surviving contact with a whole country | GE 15 cm, national | `5_deployment_finetune/` | **Every folder has its own README** giving the configs, which to run first, and the design decisions that are easy to get wrong. Start there, not here. Dataset definitions live in `configs/Custom/_base_palm/`: `dataset_uav_5cm`, `dataset_MS15_pooled`, `dataset_UAV_GE_Aerial_pooled_C`, `dataset_sat_30cm_staged`, `dataset_ge30sim`. Mapping to the internal stage names The manuscript and the authors' notes use stage letters. The folders were renamed here because `stagec` and `staged` differ by one letter and mean unrelated things, and `ms15` means nothing to a reader. | public folder | internal | |---|---| | `1_single_sensor_uav_5cm` | `maskrcnn_palm` — Stage A | | `2_pooled_15cm_ge_aerial` | `maskrcnn_palm_ms15` — Stage B | | `3_unified_multisource` | `maskrcnn_palm_stagec` — Stage C | | …