{"success":true,"data":{"author":"ECCV26-Tomato-Phenotyping authors","bucketAccessType":"public","bucketName":"ECCV26-Tomato-Phenotyping/SYNTOM","bucketPrefix":"","category":"Agriculture & Environment","cloudProvider":"huggingface","createdAt":"2026-11-21T22:15:35.556Z","description":"Procedurally generated, photorealistic renders of greenhouse tomato crops for training plant phenotyping segmentation models, released with Text-conditioned Segmentation for Tomato Phenotyping via Procedural Synthetic Data (arXiv 2607.18576), where the frames were used to fine-tune SAM 3 for tomato organs. There are 77,217 RGBA frames at 1920 x 1080 (68,328 train, 8,889 validation). Each frame has a renderer-exact semantic mask with five classes (background, leaf, stem, flower, fruit) and COCO instance annotations of whole plants, about 14.1 million instances in total, averaging 185 plants per frame and up to 849, so occlusion is heavy and thin stems and distant plants are fully labeled. Data are WebDataset tar shards under data/train/ (140 shards) and data/val/ (19 shards), each frame stored as a .png render, a .mask.png uint8 label image and a .json COCO record; annotations/instances_val.json, split lists, classes.json and dataset_stats.json sit at the root. The repository is about 356 GB (val alone about 42 GB); stream it with the webdataset library. Useful for tomato fruit and leaf segmentation, plant counting and sim-to-real transfer in greenhouse horticulture. The dataset is licensed CC BY 4.0 by its authors (the released SAM 3 weights are not); cite the paper.","displayTitle":"SYNTOM: 77,217 Synthetic Greenhouse Tomato Plant Renders with Organ Segmentation Masks and Plant Instances for Phenotyping","huggingFaceUrl":"https://huggingface.co/datasets/ECCV26-Tomato-Phenotyping/SYNTOM","id":"83a52698-c5f9-550a-a8d1-eaf4ff3d8447","license":"CC-BY-4.0","licenseEvidenceUrl":"https://huggingface.co/datasets/ECCV26-Tomato-Phenotyping/SYNTOM","objectCountEstimate":168,"probeStatus":"ok","referenceUrl":"https://huggingface.co/datasets/ECCV26-Tomato-Phenotyping/SYNTOM","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"ECCV26-Tomato-Phenotyping/SYNTOM\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"ECCV26-Tomato-Phenotyping/SYNTOM\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=ECCV26-Tomato-Phenotyping, license:cc-by-4.0, size_categories:10K<n<100K, downloads=1040"],"format":"mixed","generatedAt":"2026-11-21T22:31:39.195Z","layout":{"example":"data/train/train-000000-of-000140.tar","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"747261696e5f3033363437392e706e6700000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000030303030363434003030303030303000303030303030300030303031373234313234360030303030303030303030300030313034343500203000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000757374617200303000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000","encoding":"hex","key":"data/train/train-000000-of-000140.tar","truncatedAt":512}],"sampleObjects":["annotations/instances_val.json","classes.json","data/train/train-000000-of-000140.tar","data/train/train-000001-of-000140.tar","data/train/train-000002-of-000140.tar"],"sidecars":[{"bytes":"---\nlicense: cc-by-4.0\npretty_name: SYNTOM\nannotations_creators:\n  - machine-generated\nlanguage_creators:\n  - machine-generated\nsource_datasets:\n  - original\ntask_categories:\n  - image-segmentation\n  - object-detection\ntask_ids:\n  - semantic-segmentation\n  - instance-segmentation\nsize_categories:\n  - 10K<n<100K\ntags:\n  - agriculture\n  - agritech\n  - plant-phenotyping\n  - synthetic-data\n  - tomato\n  - greenhouse\n  - coco\n  - dense-instance-segmentation\n  - unreal-engine\n  - procedural-generation\n  - sim-to-real\n  - sam\n---\n\n# SYNTOM: Synthetic Tomato Greenhouse Segmentation\n\n![label preview](preview/contact_sheet.png)\n\n77,217 photorealistic renders of greenhouse tomato plants (68,328 train / 8,889 val,\n1920x1080) with pixel perfect ground truth for two tasks:\n\n- **Semantic segmentation**: 4 organ classes plus background, single channel PNG masks\n- **Instance segmentation**: whole plant instances in COCO format\n\nReleased with [Text-conditioned Segmentation for Tomato Phenotyping via Procedural\nSynthetic Data](https://arxiv.org/abs/2607.18576), where it is used to fine-tune SAM 3\nfor greenhouse crop organs.\n\n## Fine-tuned weights\n\nThe SAM 3 model trained on this dataset is released alongside it:\n\n**[ECCV26-Tomato-Phenotyping/SYNTOM-SAM3](https://huggingface.co/ECCV26-Tomato-Phenotyping/SYNTOM-SAM3)**\n\nIt reaches 0.6562 macro fruit IoU on three real tomato datasets against 0.5269\nfor zero-shot SAM 3, trained only on the synthetic frames published here.\n\nThe weights are a derivative of Meta's SAM 3 and carry the SAM License, not the\nCC BY 4.0 licence that covers this dataset.\n\nLabels are generated by the renderer rather than drawn by annotators, so every mask is\nexact and complete, including thin stems, occluded fruit and distant plants. Frames\naverage 185 plants and reach 849, which makes this a dense and heavily occluded\nbenchmark.\n\n## At a glance\n\n| | |\n|---|---|\n| frames | 77,217 (68,328 train / 8,889 val) |\n| resolution | 1920x1080 |\n| semantic classes | 5 (background, leaf, stem, flower, fruit) |\n| instance annotations | 14,129,160 (12,623,392 train / 1,505,768 val) |\n| plants per frame | train 185 mean / 849 max, val 169 / 835 |\n| image format | RGBA PNG |\n| label format | uint8 single channel PNG, pixel value = class id |\n| download size | 331 GiB total, 42 GiB for val alone |\n\n## Layout\n\nThe release is packed as **WebDataset** tar shards. One shard holds a few hundred\nframes, and each frame is three members sharing a key:\n\n```\nSYNTOM/\n├── data/\n│   ├── train/  train-000000-of-000140.tar ...   140 shards, 289.2 GiB\n│   └── val/    val-000000-of-000019.tar ...      19 shards, 38.1 GiB\n├── annotations/\n│   └── instances_val.json               COCO val instances, pycocotools ready\n├── splits/train.txt, splits/val.txt     one stem per line\n├── classes.json                         class ids, names and palette\n├── dataset_stats.json                   counts and per-class pixel statistics\n├── preview/contact_sheet.png            rendered label examples\n└── visualize_labels.py                  colorize and overlay the masks\n```\n\nInside a shard:\n\n```\ntrain_012345.png        the raw render, 1920x1080 RGBA\ntrain_012345.mask.png   the semantic mask, uint8, pixel value = class id\ntrain_012345.json       {\"image\": <coco image record>, \"annotations\": [...]}\n```\n\nFrames are shuffled at a fixed seed before being assigned to shards, so any single\nshard is a diverse sample rather than one contiguous camera run.\n\n## Loading\n\n```python\nfrom datasets import load_dataset\n\nds = load_dataset(\"ECCV26-Tomato-Phenotyping/SYNTOM\", split=\"val\", streaming=True)\nsample = next(iter(ds))\nimage = sample[\"png\"]              \nmask = sample[\"mask.png\"]          \nanns = sample[\"json\"][\"annotations\"]\n```\n\nOr with the `webdataset` library directly:\n\n```python\nimport io, json\nimport numpy as np\nimport webdataset as wds\nfrom PIL import Image\n\nurl = (\"https://huggingface.co/datasets/ECCV26-Tomato-Phenotyping/SYNTOM/\"\n       \"resolve/main/data/val/val-{000000..000018}-of-000019.tar\")\n\ndef decode(s","encoding":"utf8","key":"README.md","name":"README.md","size":9935,"truncated":true}],"sourceObjectKey":"data/train/train-000000-of-000140.tar","starterCell":"from huggingface_hub import hf_hub_download\npath = hf_hub_download(repo_id=\"ECCV26-Tomato-Phenotyping/SYNTOM\", filename=\"data/train/train-000000-of-000140.tar\", repo_type=\"dataset\")\nprint(path)","starterSamplePath":"annotations/instances_val.json","totalObjects":168},"summary":"Procedurally generated, photorealistic renders of greenhouse tomato crops for training plant phenotyping segmentation models, released with Text-conditioned Segmentation for…","title":"SYNTOM: 77,217 Synthetic Greenhouse Tomato Plant Renders with Organ Segmentation Masks and Plant Instances for Phenotyping","updatedAt":"2026-11-22T00:57:49.492Z","visibility":"public","websiteUrl":"https://huggingface.co/datasets/ECCV26-Tomato-Phenotyping/SYNTOM"}}