NEPEM-UFSC/dl_models

Pre-trained Neural Models for pliman

This repository hosts pre-trained Deep Learning models in ONNX and PyTorch (.pt) formats used by the pliman (Plant Image Analysis) R package for:

• Background removal & salient object segmentation;
• Open-vocabulary object detection & zero-shot instance segmentation;
• Monocular 3D depth estimation;
• Self-supervised feature extraction (DINOv2);
• Star-convex polygon detection for seeds, cells, and grain counting;
• Generative 4× super-resolution;
• Real-time YOLO object detection, instance segmentation, pose estimation, and classification.

---

📦 Available Models

Foundation & Specialized Models

Model	Filename	Size	Input Size	Task / Architecture	Description
u2netp	u2netp.onnx	4.4 MB	320×320	Salient Object Detection (U2-Net Portable)	Ultra-lightweight and fast, ideal for quick runs on CPU.
silueta	silueta.onnx	42.1 MB	320×320	Compact Matting (U2-Net variant)	High accuracy with low latency for edge & CPU execution.
u2net	u2net.onnx	167.8 MB	320×320	Salient Object Detection (U2-Net Full)	Classic robust multi-scale salient object segmentation.
rmbg-1.4	rmbg-1.4.onnx	168.0 MB	1024×1024	Background Removal (BRIA AI RMBG 1.4)	High-accuracy foreground segmentation on 1024×1024 images.
isnet-general-use	isnet-general-use.onnx	170.4 MB	1024×1024	Boundary Matting (DIS / IS-Net)	Crisp contours; excellent for complex leaf borders and lesions.
birefnet-lite	birefnet-lite.onnx	213.6 MB	1024×1024	Dichotomous Image Segmentation (BiRefNet Lite)	High-resolution bilateral reference network for fine structures.
ben2	ben2.onnx	212.6 MB	1024×1024	Boundary-Aware Extraction (BEN2 Base)	Crisp sub-pixel boundaries; excellent on leaves, seeds and roots.
withoutbg	withoutbg.onnx	433.4 MB	448×448	Depth-Aware Matting (ConvNeXt + DepthAnything)	Robust open-weights matting with depth guidance.
sam2.1 (encoder)	sam2.1.encoder.onnx	104.4 MB	1024×1024	Foundation Segmentation (Meta AI SAM 2.1 Hiera-Tiny)	Image feature extraction backbone for promptable segmentation.
sam2.1 (decoder)	sam2.1.decoder.onnx	15.8 MB	Dynamic	Mask Decoder (Meta AI SAM 2.1)	Generates crisp instance masks from bounding box prompts.
grounding-dino	groundingdino-tiny.onnx	194.4 MB	800×800	Open-Vocabulary Object Detection (Grounding DINO)	Text-prompted zero-shot bounding box detector.
vocab	vocab.txt	0.2 MB	—	BERT Vocabulary	Tokenizer vocabulary used for Grounding DINO text prompts.
sam3.1	sam3.1.onnx	868.1 MB	1024×1024	Concept-driven Segmentation (SAM 3.1)	Next-generation foundation image segmentation model.
rmbg-2.0	rmbg-2.0.onnx	976.9 MB	1024×1024	Next-Gen Matting (BRIA AI RMBG 2.0 / BiRefNet)	State-of-the-art background removal for high-detail inputs.
depth-anything-v2	depth-anything-v2-small.onnx	94.5 MB	518×518	Monocular Depth Estimation (Depth Anything V2 Small)	Dense relative 3D depth estimation from a single image.
dinov2	dinov2-vits14.onnx	84.4 MB	518×518	Vision Foundation Model (Meta AI DINOv2 ViT-S/14)	Self-supervised dense patch embeddings & semantic PCA mapping.
stardist	stardist-dsb2018.onnx	5.6 MB	256×256	Star-Convex Object Detection (StarDist DSB 2018)	Star-convex polygon detection for round/overlapping objects, seeds, and cells.
realesrgan-compact	realesrgan-compact.onnx	4.6 MB	256×256	Generative 4× Super-Resolution (Real-ESRGAN Compact)	Fast 4× image upscaling with edge preservation and tiled processing.

YOLO26 Suite (Ultralytics & ONNX)

Model	Filename	Size	Input Size	Task / Architecture	Description
yolo26n	yolo26n.onnx	9.5 MB	640×640	Real-Time Object Detection (YOLO26 Nano)	End-to-end 80 COCO classes object detection; ultra-fast CPU inference.
yolo26s	yolo26s.onnx	36.5 MB	640×640	Real-Time Object Detection (YOLO26 Small)	Balanced real-time object detector.
yolo26m	yolo26m.onnx	78.2 MB	640×640	Real-Time Object Detection (YOLO26 Medium)	Medium-size high-accuracy object detector.
yolo26l	yolo26l.onnx	95.0 MB	640×640	Real-Time Object Detection (YOLO26 Large)	High-capacity object detection.
yolo26x	yolo26x.onnx	212.9 MB	640×640	Real-Time Object Detection (YOLO26 XLarge)	Maximum precision object detector for complex scenes.
yolo26n-seg	yolo26n-seg.onnx	10.7 MB	640×640	Real-Time Instance Segmentation (YOLO26 Nano Seg)	Fast instance segmentation with continuous bilinear mask interpolation.
yolo26s-seg	yolo26s-seg.onnx	40.0 MB	640×640	Real-Time Instance Segmentation (YOLO26 Small Seg)	Balanced instance segmentation.
yolo26m-seg	yolo26m-seg.onnx	90.2 MB	640×640	Real-Time Instance Segmentation (YOLO26 Medium Seg)	High-precision instance masks and morphological metrics.
yolo26l-seg	yolo26l-seg.onnx	107.1 MB	640×640	Real-Time Instance Segmentation (YOLO26 Large Seg)	Large-capacity instance segmentation.
yolo26x-seg	yolo26x-seg.onnx	240.0 MB	640×640	Real-Time Instance Segmentation (YOLO26 XLarge Seg)	Highest fidelity instance segmentation and boundary precision.
yolo26n-pose	yolo26n-pose.onnx	11.6 MB	640×640	Real-Time Pose Estimation (YOLO26 Nano Pose)	17 COCO anatomical keypoints and skeleton limb connections.
yolo26s-pose	yolo26s-pose.onnx	39.9 MB	640×640	Real-Time Pose Estimation (YOLO26 Small Pose)	Balanced speed/precision keypoint detection.
yolo26m-pose	yolo26m-pose.onnx	82.6 MB	640×640	Real-Time Pose Estimation (YOLO26 Medium Pose)	Medium pose estimation model.
yolo26l-pose	yolo26l-pose.onnx	99.4 MB	640×640	Real-Time Pose Estimation (YOLO26 Large Pose)	Large-capacity pose estimation model.
yolo26x-pose	yolo26x-pose.onnx	220.0 MB	640×640	Real-Time Pose Estimation (YOLO26 XLarge Pose)	Maximum accuracy pose estimation for challenging poses and occlusion.
yolo26n-cls	yolo26n-cls.onnx	10.8 MB	224×224	Image Classification (YOLO26 Nano Cls)	Fast 1,000-class ImageNet classification with Softmax probabilities.
yolo26s-cls	yolo26s-cls.onnx	25.7 MB	224×224	Image Classification (YOLO26 Small Cls)	Balanced ImageNet classification.
yolo26m-cls	yolo26m-cls.onnx	44.4 MB	224×224	Image Classification (YOLO26 Medium Cls)	High accuracy ImageNet classification.
yolo26l-cls	yolo26l-cls.onnx	53.9 MB	224×224	Image Classification (YOLO26 Large Cls)	Large ImageNet classification model.

---

🚀 Usage in pliman

In R, models can be downloaded automatically with live CLI progress bars:

library(pliman)
# Check available models and their download status
pliman_available_models()
# Download a model (e.g. u2net)
pliman_download_model("u2net")
# Or download all available models at once
pliman_download_model("all")
img <- image_import("leaves.jpg")
# 1. Background removal / foreground segmentation
seg <- image_segment_dl(img, model = "u2netp")
# 2. Text-prompted zero-shot instance segmentation (Grounded-SAM)
res <- image_segment_dl(
  img,
  model = "grounded-sam",
  prompt = "leaf",
  type = "highlight",
  bbox = TRUE
)
# 3. Real-time object detection (YOLO26)
boxes <- image_detect_dl(img, model = "yolo26n")
# 4. Real-time instance segmentation (YOLO26-seg)
yolo_seg <- image_segment_dl(img, model = "yolo26n-seg", type = "highlight")
# 5. Real-time human pose estimation (YOLO26-pose)
pose_res <- image_pose_dl(img, model = "yolo26n-pose")
# 6. Image classification (YOLO26-cls, Top-5 ImageNet classes)
top_cls <- image_classify_dl(img, model = "yolo26n-cls", top_k = 5)
# 7. Star-convex polygon detection for seeds, grains, and cells (StarDist)
stars <- image_stardist_dl(img, type = "segment")
# 8. Monocular 3D depth estimation (Depth Anything V2)
depth <- image_depth_dl(img, col_palette = "viridis")
# 9. Self-supervised feature extraction and semantic PCA mapping (DINOv2)
feats <- image_features_dl(img)
# 10. 4x Generative super-resolution (Real-ESRGAN Compact)
sr_img <- image_superres_dl(img, scale = 4)

📜 Acknowledgments & References

These models and weights originate from foundational research and open-source contributions across the computer vision community:

• U2-Net / U2-Netp: Qin, X., Zhang, Z., Huang, C., Dehghan, M., Zaiane, O. R., & Jagersand, M. (2020). U2-Net: Going deeper with nested U-structure for salient object detection. Pattern Recognition, 106, 107404. Paper | GitHub
• IS-Net / DIS: Xue, P., Colin, P. A., Qin, X., & Jagersand, M. (2022). Highly accurate dichotomous image segmentation. In European Conference on Computer Vision (ECCV). Paper | GitHub
• BiRefNet: Zheng, P., Gao, D., Fan, D. P., Cheng, M. M., & Van Gool, L. (2024). Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research. Paper | GitHub
• RMBG 1.4 & 2.0: BRIA AI. RMBG: State-of-the-Art Background Removal Model. Hugging Face RMBG 1.4 | Hugging Face RMBG 2.0
• BEN2: Prama LLC. Boundary-aware Extraction Network (BEN2). GitHub
• withoutBG: withoutBG. Depth-Guided Salient Matting with ConvNeXt & DepthAnything. GitHub
• Segment Anything 2.1 (SAM 2.1): Ravi, N., Gabeur, V., Hu, Y. T., Hu, R., Ryali, C., Ma, T., ... & Feichtenhofer, C. (2024). SAM 2: Segment Anything in Images and Videos. Meta AI. Paper | GitHub
• Grounding DINO: Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., ... & Zhang, L. (2023). Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. In European Conference on Computer Vision. Paper | GitHub
• Depth Anything V2: Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., & Zhao, H. (2024). Depth Anything V2: A More Capable Foundation Model for Monocular Depth Estimation. Paper | GitHub
• DINOv2: Oquab, M., Darcet, T., Moutakanni, T., Vo, H. V., Szafraniec, M., Khalidov, V., ... & Bojanowski, P. (2024). DINOv2: Learning Robust Visual Features without Supervision. Transactions on Machine Learning Research (TMLR). Paper | GitHub
• YOLO26 & Ultralytics: Ultralytics. YOLO26: Next-Generation Real-Time Object Detection, Instance Segmentation, Pose Estimation, and Classification. Ultralytics | Hugging Face Hub
• StarDist: Schmidt, U., Weigert, M., Broaddus, C., & Myers, G. (2018). Cell Detection with Star-Convex Polygons. In Medical Image Computing and Computer Assisted Intervention (MICCAI), pp. 265–273. Paper | GitHub
• Real-ESRGAN Compact: Wang, X., Xie, L., Dong, C., & Shan, Y. (2021). Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data. In IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). Paper | GitHub