{"success":true,"data":{"author":"SelvaBox / CanopyRS authors","bucketAccessType":"public","bucketName":"CanopyRS/SelvaBox","bucketPrefix":"","category":"Environmental Science","cloudProvider":"huggingface","createdAt":"2026-11-21T22:15:35.735Z","description":"A high-resolution benchmark for individual tree crown detection in dense tropical rainforest canopies, introduced in the SelvaBox paper (ICLR 2026, arXiv 2507.00170) and used to train and benchmark the CanopyRS tree detection and segmentation pipeline. Fourteen drone (UAV) RGB orthomosaics from Brazil, Ecuador and Panama were resampled to 4.5 cm ground sampling distance and tiled, with more than 83,000 unique human-drawn bounding boxes around tree crowns. Training tiles are 3555 x 3555 pixels (160 m) with 50 percent overlap; validation and test tiles are 1777 x 1777 pixels (80 m) with 50 and 75 percent overlap so the largest crowns fit in a tile, giving 585 train, 387 validation and 1,477 test tiles. Labels are binary tree versus background, without species. Data are Parquet shards under data/ (about 31 GB) with TIFF images and COCO-format boxes, loadable or streamable with datasets.load_dataset; GeoPackage annotations and AOIs are on the gpkg branch. Useful for forest monitoring, canopy mapping, tree counting and object detection under heavy overlap. Licensed CC BY 4.0 by the authors; cite the SelvaBox paper (arXiv 2507.00170).","displayTitle":"SelvaBox: 83,000 Tropical Tree Crown Bounding Boxes on 4.5 cm Drone Orthomosaics from Brazil, Ecuador and Panama (CanopyRS)","huggingFaceUrl":"https://huggingface.co/datasets/CanopyRS/SelvaBox","id":"7be6f6e4-cf9b-5168-9494-fdfee4919c31","license":"CC-BY-4.0","licenseEvidenceUrl":"https://huggingface.co/datasets/CanopyRS/SelvaBox","objectCountEstimate":67,"probeStatus":"ok","referenceUrl":"https://huggingface.co/datasets/CanopyRS/SelvaBox","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"CanopyRS/SelvaBox\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"CanopyRS/SelvaBox\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=CanopyRS, license:cc-by-4.0, size_categories:1K<n<10K, format:parquet, downloads=675"],"format":"mixed","generatedAt":"2026-11-21T22:31:42.662Z","layout":{"example":"data/test-00000-of-00024.parquet","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"data/test-00000-of-00024.parquet","truncatedAt":512}],"sampleObjects":["data/test-00000-of-00024.parquet","data/test-00001-of-00024.parquet","data/test-00002-of-00024.parquet","data/test-00003-of-00024.parquet","data/test-00004-of-00024.parquet"],"sidecars":[{"bytes":"---\ndataset_info:\n  features:\n  - name: height\n    dtype: int64\n  - name: width\n    dtype: int64\n  - name: fold\n    dtype: string\n  - name: raster_name\n    dtype: string\n  - name: location\n    dtype: string\n  - name: image\n    dtype: image\n  - name: tile_name\n    dtype: string\n  - name: annotations\n    struct:\n    - name: bbox\n      sequence:\n        sequence: float64\n    - name: segmentation\n      dtype: 'null'\n    - name: area\n      sequence: float64\n    - name: iscrowd\n      sequence: int64\n    - name: is_rle_format\n      dtype: 'null'\n    - name: category\n      sequence: string\n  - name: tile_metadata\n    struct:\n    - name: crs\n      dtype: string\n    - name: transform\n      sequence: float64\n    - name: bounds\n      sequence: float64\n    - name: width\n      dtype: int64\n    - name: height\n      dtype: int64\n    - name: count\n      dtype: int64\n    - name: dtypes\n      sequence: string\n    - name: nodata\n      dtype: 'null'\n  splits:\n  - name: test\n    num_bytes: 11729504363.402\n    num_examples: 1477\n  - name: validation\n    num_bytes: 2786536280\n    num_examples: 387\n  - name: train\n    num_bytes: 16884458976\n    num_examples: 585\n  download_size: 31336873232\n  dataset_size: 31400499619.402\nconfigs:\n- config_name: default\n  data_files:\n  - split: test\n    path: data/test-*\n  - split: validation\n    path: data/validation-*\n  - split: train\n    path: data/train-*\nlicense: cc-by-4.0\ntags:\n- vision\n- ai\n- climate\n- forest\n- tree\n- remote_sensing\nsize_categories:\n- 10K<n<100K\n---\n\n---\n\n# SelvaBox: A high-resolution dataset for tropical tree crown detection\n\n<p align=\"left\">\n  <img src=\"./teaser_wo_name.png\" width=\"800\"/>\n</p>\n\n<!-- Provide a quick summary of the dataset. -->\n\nThis is the version of the SelvaBox dataset that has been pre-processed and presented in our SelvaBox paper.\nThe dataset is made of 14 rasters resampled at 4.5 cm GSD, from three different countries: Brazil, Ecuador and Panama. These rasters were tiled into more than 2400 images. It comprises over 83 000 unique human bounding box annotations for tropical tree crowns in dense canopies.\n\nSelvaBox is used to train and benchmark models in [CanopyRS](https://github.com/hugobaudchon/CanopyRS), a pipeline for detecting and segmenting tree crowns in high-resolution orthomosaics ([documentation](https://hugobaudchon.github.io/CanopyRS/)).\n\n## Dataset Details\n\n### Dataset Description\n\n<!-- Provide a longer summary of what this dataset is. -->\n\nTraining tiles are 3555x3555 pixels, while validation and test tiles are 1777x1777 pixels, equivalent to 80x80 meters spatial extent. There is 50% overlap between train and validation tiles, and 75% between test tiles (to ensure that the largest trees of 50+ meters in diameter will fit entirely in at least one tile).\nThe table below summarizes the information on the three splits. Note that the # Annotations reported is larger than 83000 due to the overlap between tiles, which duplicates annotations. \nThere is also a similar effect regarding the # Tiles: there are more test tiles than train or valid but that's because of the 75% overlap between tiles, compared to 50%.\nThe 'Geographic Area % of total dataset column' more accurately describes how much data was assigned to each split.\n\n| Split | Tile Size (px) | Tile Size (m) | Overlap | # Tiles | # Annotations | Geographic Area % of total dataset|\n| ----- | -------------- | ------------- | ------- | ------- | ------------- | ------------------- |\n| Train | 3555           | 160.0 m       |    50%     | 585     | 232,071       | ~74%               |\n| Valid | 1777           | 80.0 m        |    50%     | 387     | 38,651        | ~13%                |\n| Test  | 1777           | 80.0 m        |    75%     | 1,477   | 161,188       | ~13%               |\n\n\n- **Curated by:** Will be added after double-blind review.\n- **Funded by:** Will be added after double-blind review.\n- **License:** CC BY 4.0\n\n### Dataset Sources\n\n<!-- Provide the basic links for the dataset. -->\n\n- **Repository:** [CanopyRS](https://github.com/hugobaudchon/CanopyRS)\n- **Paper:** [SelvaBox — ICLR 2026](","encoding":"utf8","key":"README.md","name":"README.md","size":12946,"truncated":true}],"sourceObjectKey":"data/test-00000-of-00024.parquet","starterCell":"from datasets import load_dataset\nds = load_dataset(\"CanopyRS/SelvaBox\")\nprint(ds)","starterSamplePath":"data/test-00000-of-00024.parquet","totalObjects":67},"summary":"A high-resolution benchmark for individual tree crown detection in dense tropical rainforest canopies, introduced in the SelvaBox paper (ICLR 2026, arXiv 2507.00170) and used to…","title":"SelvaBox: 83,000 Tropical Tree Crown Bounding Boxes on 4.5 cm Drone Orthomosaics from Brazil, Ecuador and Panama (CanopyRS)","updatedAt":"2026-11-22T00:59:35.438Z","visibility":"public","websiteUrl":"https://github.com/hugobaudchon/CanopyRS"}}