Granularity-Aware Transfer for Tree Instance Segmentation in Synthetic and Real Forests

Pankaj Deoli; Atef Tej; Anmol Ashri; Anandatirtha JS; Karsten Berns

cs.CVApr 15, 2026

Granularity-Aware Transfer for Tree Instance Segmentation in Synthetic and Real Forests

Authors: Pankaj Deoli, Atef Tej, Anmol Ashri, Anandatirtha JS, Karsten Berns

Organizations: Robotics Research Lab, University of Kaiserslautern-Landau, Kaiserslautern, Germany · University of Kaiserslautern-Landau, Kaiserslautern, Germany

Abstract

We address the challenge of synthetic-to-real transfer in forestry perception where real data have only coarse Tree labels while synthetic data provide fine-grained trunk/crown annotations. We introduce MGTD, a mixed-granularity dataset with 53k synthetic and 3.6k real images, and a four-stage protocol isolating domain shift and granularity mismatch. Our core contribution is granularity-aware distillation, which transfers structural priors from fine-grained synthetic teachers to a coarse-label student via logit-space merging and mask unification. Experiments show consistent mask AP gains, especially for small/distant trees, establishing a testbed for Sim-Real transfer under label granularity constraints.

Figures & tables

Panel of Table 1 — Table 1: Examples of RGB images, Bounding Boxes (BB), Oriented Bounding Boxes (OBB), Instance Masks (IM), and Oriented Instance Masks (OIM) for simulated (row 2,3) and real (row 1) forest scenes in the MGTD dataset.

Model	Type	Backbone	[email protected]:0.95 (BBox)	[email protected] (BBox)	[email protected]:0.95 (Segm)
Mask-RCNN	Tree trunk	X101	0.732	0.904	0.686
Mask-RCNN	Tree trunk	Swin-T	0.788	0.936	0.753
Mask-RCNN	Whole tree	X101	0.859	0.947	0.670
Mask-RCNN	Whole tree	Swin-T	0.880	0.948	0.688

Table 2: Phase 1 (synthetic only). Teacher models trained on tree trunk and whole tree simulation data, evaluated on the synthetic validation set. We observe that Swin-T outperforms X101 (for both Mask RCNN & Cascade Mask RCNN) and is therefore chosen as the final teacher. Full results are provided in the supplementary (Table 8).

Panel of Figure 1 — Figure 1: Phase 1: Instance segmentation results on synthetic validation data. Top row shows tree trunk example, bottom row shows whole tree . Columns display RGB input, ground-truth and model prediction. More results in supplementary (Fig. 2 & Fig. 3).

Model	Type	Task	Sim Val	Sim $\rightarrow$ Real	Abs. Gap (Rel. Drop)
Mask-RCNN	Tree trunk	BBox	0.788	0.225	0.563 (71.4%)
Mask-RCNN	Tree trunk	Segm	0.753	0.147	0.606 (80.5%)
Mask-RCNN	Whole tree	BBox	0.880	0.066	0.814 (92.5%)
Mask-RCNN	Whole tree	Segm	0.688	0.039	0.649 (94.3%)

Table 3: Phase 2 (Sim

\rightarrow

Real). Synthetic teachers evaluated zero-shot on the real validation set with label-agnostic AP. We report performance on synthetic vs. real data along with absolute and relative drops. We use the Swin-T backbone for the below-mentioned tests.

Panel of Figure 2 — Figure 2: Phase 2: Domain transfer. Comparison of predictions obtained from tree trunk and whole tree teachers when evaluated on real forest images.

Model	Type	Backbone	[email protected]:0.95 (BBox)	[email protected] (BBox)	[email protected]:0.95 (Segm)
Mask R-CNN	Tree	R50	0.487	0.817	0.396
Mask R-CNN	Tree	X101	0.461	0.810	0.363
Mask R-CNN	Tree	Swin-T	0.500	0.792	0.420

Table 4: Phase 3 (real-only) results. Students are evaluated on the real validation set. “BBox” reports COCO-style bounding-box AP, and “Segm” reports mask AP. [email protected]:0.95 averages AP across IoU thresholds, while [email protected] is AP at IoU = 0.5.

Panel of Figure 3 — Figure 3: Qualitative comparison of real-only training (Phase 3) and granularity-aware distillation (Phase 4). Phase 4 recovers additional thin and distant tree trunks and exhibits stronger vertical consistency, while Phase 3 often misses such instances. We also see an improvement in distinguishing capabilities (row 4 for e.g.).

Model (Backbone)	Training Regime	mAP 50:95 (Segm)	Improvement
Mask R-CNN (ResNet-50)	Phase 3 (Real-only)	0.396	–
Mask R-CNN (ResNeXt-101)	Phase 3 (Real-only)	0.363	–
Mask R-CNN (Swin-T)	Phase 3 (Real-only)	0.420	–
Mask R-CNN (ResNet-50)	Phase 4 (Distillation)	0.429	+8.3% vs. same backbone
Mask R-CNN (ResNet-50)	Phase 4 (Distillation)	0.429	Outperforms ResNeXt-101 (0.363)
			Outperforms Swin-T (0.420)

Table 5: Phase 4 (Granularity-aware distillation) vs. Phase 3 (Real-only baseline). Despite using a lighter ResNet-50 backbone, distillation surpasses all real-only models.

Subset	Train (I/O/A)	Val (I/O/A)	Test (I/O/A)	Total (I/O)
1 / Tree Trunk	10036 / 111068 / 11.07	2867 / 31832 / 11.10	1434 / 15632 / 10.90	14337 / 158532
1 / Whole Tree	8921 / 29282 / 3.28	2549 / 8338 / 3.27	1274 / 4184 / 3.28	12744 / 41704
2 / Tree Trunk	9712 / 101677 / 10.47	2775 / 29301 / 10.56	1388 / 14474 / 10.43	13875 / 145452
2 / Whole Tree	9400 / 47411 / 5.04	2686 / 13714 / 5.11	1343 / 6811 / 5.07	13429 / 67936
3 / Tree Trunk	7447 / 84651 / 11.37	2128 / 23852 / 11.21	1064 / 12314 / 11.57	10639 / 120817
3 / Whole Tree	7337 / 45580 / 6.21	2096 / 12606 / 6.01	1049 / 6486 / 6.18	10482 / 64672

Table 6: Dataset statistics for the simulated part. Each entry shows #images / #objects / average objects per image (I/O/A). Subsets 1–4 correspond to dawn, dusk, noon, and snow respectively.

Subset	Train (I/O/A)	Val (I/O/A)	Test (I/O/A)	Total (I/O)
Part 1	1400 / 14065 / 10.05	400 / 4064 / 10.16	200 / 1974 / 9.87	2000 / 20103
Part 2	1120 / 9156 / 8.18	320 / 2663 / 8.32	160 / 1285 / 8.03	1600 / 13104

Table 7: Dataset statistics for the real part of the dataset, divided into two subsets. Each entry shows #images / #objects / average objects per image (I/O/A).

Panel of Figure 4 — Figure 4: Sample images from the MGTD dataset. Includes both the real and simulated part.

Parameter Group	Setting
Image size	1024 $\times$ 1024 (fixed), multi_scale = false
Batching	Batch size = 1 (auto-batch), single GPU; no gradient accumulation
Optimizer	SGD (momentum = 0.937, weight decay = 5e–4)
Learning rate schedule	Linear LR; base LR = 0.01; final LR fraction = 0.01 $\rightarrow$ final LR = $1\times 10^{-4}$ ; Warmup: 3 epochs (momentum = 0.8, bias_lr = 0.0)
Augmentations	Mosaic: on (prob = 1.0), disabled in last 10 epochs; MixUp: off; HSV jitter: h = 0.015, s = 0.7, v = 0.4; RandomAffine: translate = 0.1, scale = 0.5; Flips: fliplr = 0.5, flipud = 0.0; Random Erasing: prob = 0.4; CutMix / Copy-Paste: off; AutoAugment: RandAugment
Losses	Box: 7.5 (CIoU style); Class: 0.5 (BCE); DFL: 1.5 (distributional regression); Label smoothing: none

Table 8: Hyperparameters for YOLOv8 / YOLOv11 (n, s, m, l, x) on simulated trees for both axis-aligned and oriented bounding boxes. 2 2 2 Simulated dataset trained for 100 epochs, real dataset trained for 50 epochs due to higher overfitting risk and greater per-iteration computational cost.

Parameter Group	Setting
Image size	640 $\times$ 640 (fixed, multi_scale: false)
Batching	Batch size = 16, single/multi-GPU depending on device, no gradient accumulation
Optimizer	SGD (momentum: 0.937, weight decay: 0.0005)
Learning rate schedule	Linear LR (cos_lr: false), base LR = 0.01, final LR fraction = 0.01 $\rightarrow$ ends at $1\times 10^{-4}$ ; warmup (epochs = 3.0, momentum = 0.8, bias_lr = 0.1)
Augmentations	Same as simulated trees: Mosaic on (prob = 1.0), off in last 10 epochs (close_mosaic = 10); MixUp off; HSV jitter (h = 0.015, s = 0.7, v = 0.4); RandomAffine (degrees = 0.0, translate = 0.1, scale = 0.5, shear = 0.0, perspective = 0.0); Flips (flipud = 0.0, fliplr = 0.5); CutMix off; Random Erasing on (prob = 0.4); Copy-Paste off; AutoAugment: RandAugment
Losses	Box loss: 7.5 (CIoU-style); Class loss: 0.5 (BCE, no focal); DFL: 1.5 (distributional regression on); Label smoothing: 0.0

Table 9: Hyperparameters for YOLOv8 / YOLOv11 (n, s, m, l, x) – Real Trees † † footnotemark:

Parameter Group	Setting
Backbone & Init	RegNet (return_idx = [1,2,3]); initialized with ImageNet-1K weights (torchvision)
Epochs	130 total
Optimizer	AdamW; LR = 1e-4 (all modules); weight decay = 1e-4 (bias/norm excluded); gradient clipping = 0.1
LR Schedule	MultiStepLR (milestones = [1000], $\gamma=0.1$ → no effective decay); no warmup
Queries & Matcher	300 object queries; Hungarian costs — cls: 2, L1: 5, GIoU: 2; focal matcher $(\alpha=0.25,\gamma=2.0)$
Losses	Varifocal classification loss $(\alpha=0.75,\gamma=2.0)$ ; Box: L1 (w=5) + GIoU (w=2); Denoising with 100 queries

Table 10: Hyperparameter setup for RT-DETR used in our experiments.

Model	Backbone	Optimizer	LR	Weight Decay	Schedule / Augmentations
Mask R-CNN	ResNeXt-101-64x4d + FPN	SGD ( $m=0.9$ )	0.02	0.0001	Step LR + linear warmup ( $r=0.001$ ), Resize + Flip + Norm + Pad
Mask R-CNN	Swin-T + FPN	AdamW	1e-4	0.05	Step LR + linear warmup ( $r=0.001$ ), Resize + Flip + Norm + Pad
Cascade Mask R-CNN	ResNeXt-101-64x4d + FPN	SGD ( $m=0.9$ )	0.02	0.0001	Step LR + linear warmup ( $r=0.001$ ), Resize + Flip + Norm + Pad
Cascade Mask R-CNN	Swin-T + FPN	AdamW	1e-4	0.05	Step LR + linear warmup ( $r=0.001$ ), Resize + Flip + Norm + Pad
Mask2Former	R50_lsj_50e	AdamW	1e-4	0.05	Step LR + linear warmup ( $r=0.001$ ), Resize + Flip + Norm + Pad
Mask2Former	R101_lsj_50e / tuned	AdamW	1e-4	0.05	Step LR + linear warmup ( $r=0.001$ ), Resize + Flip + Norm + Pad

Table 11: Hyperparameter setup for instance segmentation on real and simulated trees (MMDetection). Models are grouped by family, repeated values are merged for compactness.

Hyperparameter	Value
Teacher architecture	Mask R-CNN (Swin-T, MMDetection).
Student architecture	Mask R-CNN (ResNet-50 + FPN).
Training epochs	100
Batch size (total)	48
Labeled batch size	16
Data loader workers	8

Table 12: Training hyperparameters for Phase 4 (granularity-aware distillation).

Type	Model	Backbone	Epochs	mAP@[.5:.95] (Box)	mAP@50 (Box)	mAP@75 (Box)	APs (Box)	APm (Box)	APl (Box)	mAR@100	mAR@300	mAR@1000	ARs	ARm	ARl
Tree Trunk	Mask-RCNN	X101	24	0.732	0.904	0.858	0.694	0.769	0.837	0.812	0.812	0.812	0.778	0.840	0.916
		Swin-T	24	0.788	0.936	0.908	0.772	0.801	0.868	0.847	0.847	0.847	0.822	0.865	0.939
	Cascade Mask-RCNN	X101	24	0.762	0.905	0.872	0.728	0.801	0.877	0.829	0.829	0.829	0.797	0.858	0.937
		Swin-T	24	0.764	0.912	0.877	0.729	0.803	0.874	0.837	0.837	0.837	0.798	0.844	0.942
Whole Tree	Mask-RCNN	X101	24	0.859	0.947	0.930	0.717	0.880	0.856	0.905	0.905	0.905	0.764	0.895	0.909
		Swin-T	24	0.880	0.948	0.936	0.798	0.903	0.874	0.923	0.923	0.923	0.823	0.916	0.926

Table 13: Phase 1: Bounding-box (top) and segmentation (bottom) results on simulated datasets (tree trunk and whole tree). Metrics include AP/AR at multiple IoU thresholds and object scales. Bold numbers highlight best values in each block.

Panel of Figure 5 — Figure 5: Phase 1: Instance segmentation results on Tree Trunk examples. Each row shows the RGB image, model prediction, and ground truth mask.

Panel of Figure 6 — Figure 6: Phase 1: Instance segmentation results on Whole tree examples. Each row shows the RGB image, model prediction, and ground truth mask.

Panel of Figure 7 — Figure 7: Phase 2: Domain transfer (Tree Trunk $\rightarrow$ Real Trees).

Model	Type	Epochs	[email protected]:0.95	[email protected]	[email protected]	APs	APm	API	mAR@100	AR@300	AR@1000	ARI
Bounding Box Metrics
Mask-RCNN (Swin-T)	Tree trunk	24	0.225	0.445	0.205	-1.000	-1.000	0.225	0.364	0.364	0.364	0.364
Mask-RCNN (Swin-T)	Whole tree	24	0.066	0.191	0.038	-1.000	-1.000	0.066	0.178	0.178	0.178	0.178
Segmentation Metrics
Mask-RCNN (Swin-T)	Tree trunk	24	0.147	0.300	0.127	-1.000	-1.000	0.147	0.249	0.249	0.249	0.249
Mask-RCNN (Swin-T)	Whole tree	24	0.039	0.104	0.024	-1.000	-1.000	0.039	0.131	0.131	0.131	0.131

Table 14: Phase 2: Validation metrics for direct Sim

\rightarrow

Real transfer on the validation set. Bounding box and segmentation metrics are separated by a dashed line.

Panel of Figure 8 — Figure 8: Phase 2: Domain transfer (Whole Tree $\rightarrow$ Real Trees).

Task	Model	Backbone	Epochs	mAP@[.5:.95]	mAP@50	AP50	AP75	APs	APm	APl	AR@[.5:.95]
Bbox	Mask-RCNN	Swin-T	24	0.500	0.792	0.792	0.544	-	-	0.500	0.617
		X-101	24	0.461	0.810	0.810	0.467	-	-	0.461	0.592
	Cascade RCNN	X-101	24	0.571	0.834	0.834	0.629	-	-	0.571	0.684
		Swin-T	24	0.571	0.842	0.842	0.629	-	-	0.571	0.685
	Mask2Former	R50	24	0.265	0.623	0.623	0.186	-	-	0.265	0.510
		R101	24	0.280	0.662	0.662	0.194	-	-	0.280	0.523

Table 15: Phase 3: Instance segmentation results on real tree dataset. For Mask2Former, validation results were not fully available; test results are reported for completeness. The test results showed similar pattern as validation.

Panel of Figure 9 — Figure 9: Phase 3: Instance segmentation (Mask-RCNN with Swin-T backbone) results on real trees (trained directly on real data). Each row shows RGB image, prediction, and ground truth.

Metric	Tree Trunk + Student	Whole Tree + Student	Key Insight
Final IoU	0.512	0.630	Whole tree supervision leads to better final overlap .
Peak IoU	0.689 (Epoch 97)	0.667 (Epoch 90)	Trunk supervision enables higher best-ever overlap.
Final mAP@[.5:.95]	0.138	0.326	Whole tree supervision yields better final detection quality .
Peak mAP@[.5:.95]	0.427 (Epoch 88)	0.381 (Epoch 90)	Trunk supervision achieves superior peak detection precision .
Stability (Post-Epoch 50)	High	Low	Trunk teacher provides more consistent learning signals .

Table 16: Final and peak performance metrics for student models trained with individual teachers. The tree trunk teacher facilitates higher peak precision (mAP), while the whole tree teacher contributes to final overlap accuracy (IoU).

Panel of Figure 10 — Figure 10: Phase 1: Qualitative analysis of yolov11m (trained on simulated tree trunks) on the simulated tree trunks val set.

Panel of Figure 11 — Figure 11: Phase 1: Qualitative analysis of yolov8m (trained on simulated whole trees) on the simulated whole trees val set.

Task	Model	Epoch	train/box_loss	train/cls_loss	train/dfl_loss	Precision	Recall	mAP@50	mAP@50:95	val/box_loss	val/cls_loss	val/dfl_loss	lr/pg0	lr/pg1	lr/pg2
Tree Trunk	YOLOv8n	100	0.64198	0.51119	0.85470	0.82755	0.87099	0.92217	0.77399	0.64272	0.53417	0.85107	0.000199	0.000199	0.000199
	YOLOv8s	100	0.55581	0.42323	0.83554	0.83738	0.87747	0.92875	0.79685	0.59484	0.49071	0.84336	0.000199	0.000199	0.000199
	YOLOv8m	100	0.52055	0.39357	0.83282	0.83614	0.88321	0.93024	0.80409	0.57936	0.47443	0.84955	0.000199	0.000199	0.000199
	YOLOv8l	100	0.50205	0.36515	0.82760	0.84004	0.88467	0.93074	0.80960	0.56565	0.46984	0.85554	0.000199	0.000199	0.000199
	YOLOv11n	100	0.62017	0.50105	0.84677	0.83320	0.87276	0.92494	0.78447	0.61998	0.52759	0.84075	0.000199	0.000199	0.000199
	YOLOv11s	100	0.54407	0.42742	0.83043	0.83856	0.88160	0.93020	0.80234	0.58274	0.48856	0.83403	0.000199	0.000199	0.000199

Table 17: Phase 1: Validation metrics for object detection on simulated tree trunk and whole tree datasets. Models are ordered by variant size (n, s, m, l).

Type (Tree trunk / Whole tree)	Model	[email protected]:0.95	[email protected]	AP50	AP75	APs	APm	API	mAR@100
Whole tree	RT-DETR + RegNet	0.920	0.962	0.962	0.957	0.817	0.938	0.917	0.974
Tree trunk	RT-DETR + RegNet	0.825	0.947	0.947	0.928	0.797	0.861	0.899	0.973

Table 18: Phase 1: Validation metrics for object detection on the validation set (tree trunk vs whole tree) using RT-DETR + RegNet.

Figure 12: Phase 2: Domain gap when the best simulated (tree trunk and whole tree model) was directly tested on real tree images.

Panel of Table 19 — Table 19: Qualitative analysis of YOLOv11m (trained on simulated tree trunks) on the real trees test set.

Panel of Table 20 — Table 20: Qualitative analysis of YOLOv8m (trained on simulated whole trees) on the real trees test set.

Model	Type	Task	Simulated accuracy (val set)	Sim→Real acc.	Absolute Gap	Relative Drop
yolo11m	Tree trunk	BBox	0.93341	0.40	0.53	56.8%
yolov8m	Whole tree	BBox	0.724	0.4817	0.24	33%

Table 21: Phase 2 : Domain gap analysis between simulated and real validation performance. YOLO series

Task	Model	[email protected]:0.95	[email protected]	AP50	AP75	APs	APm	API	Pretrained on
Object detection	RT-DETR + RegNet	0.232	0.438	0.438	0.220	-1	-1	0.232	Tree trunk sim
Object detection	RT-DETR + RegNet	0.062	0.182	0.182	0.036	-1	-1	0.062	Whole tree sim

Table 22: Phase 2: Domain gap analysis between simulated and real validation performance. (RT-DETR + RegNet).

Task	Model	Epoch	train/box_loss	train/cls_loss	train/dfl_loss	Precision	Recall	mAP@50	mAP@50:95	val/box_loss	val/cls_loss	val/dfl_loss	lr
Object Detection	YOLOv8n	50	0.9830	0.8459	1.0656	0.821	0.805	0.879	0.616	0.928	0.837	1.058	9.92e-05
	YOLOv8s	50	0.8869	0.7426	1.0244	0.816	0.826	0.886	0.637	0.928	0.772	1.058	9.92e-05
	YOLOv8m	50	0.8173	0.6936	1.0125	0.821	0.825	0.886	0.649	0.898	0.720	1.120	9.92e-05
	YOLOv8l	50	0.8129	0.6565	1.0225	0.828	0.823	0.890	0.656	0.882	0.705	1.126	9.92e-05
	YOLOv8x	50	0.7900	0.6260	1.0095	0.818	0.835	0.892	0.660	0.883	0.690	1.140	9.92e-05
	YOLOv11s	50	0.9979	0.8583	1.0650	0.787	0.807	0.863	0.590	0.961	0.814	1.102	5.96e-05

Table 23: Phase 3: Validation metrics for object detection on real trees across YOLOv8 and YOLOv11 variants. All models were trained for 50 epochs. (YOLO series)

Task	Model	[email protected]:0.95	[email protected]	AP50	AP75	APs	APm	API
Object detection	RT-DETR + RegNet	34.7	68.6	68.6	30.3	-1	-1	34.7

Table 24: Phase 3: Validation metrics for RT-DETR + RegNet trained on real trees.

Instance Segmentation

Instance segmentation identifies each object and assigns it a separate pixel-level mask in images or video.

See all 159 papers

FD $^2$ : A Dedicated Framework for Fine-Grained Dataset Distillation
Mar 26, 2026Hongxu Ma, Guang Li, Shijie Wang +5Fine-Grained Image Classification Dataset Distillation
Zhejiang University · Hokkaido University · The University of Queensland +2
Cross-Domain Dead Tree Detection via Knowledge Distillation in Aerial Imagery
Jun 1, 2026Anis Ur Rahman, Mete Ahishali, Einari Heinaro +1Remote Sensing Image Understanding Cross-Domain Transfer Learning
CSC – IT Center for Science Ltd., Espoo, Finland · Department of Forest Sciences, University of Helsinki · KOKO Forest Ltd. +1
SegmentAnyTreeV2: Scaling Transformer-Based Tree Instance Segmentation Across Sensors, Platforms, and Forests
Jun 6, 2026Maciej Wielgosz, Stefano Puliti, Rasmus AstrupTransformer Instance Segmentation
Norwegian Institute of Bioeconomy Research (NIBIO), Høgskoleveien 7, 1433 Ås, Norway

Mar 26, 2026J/K move · Enter open · S save

FD $^2$ : A Dedicated Framework for Fine-Grained Dataset Distillation

Hongxu Ma, Guang Li, Shijie Wang, Dongzhan Zhou+4

Zhejiang University · Hokkaido University · The University of Queensland+2

Fine-Grained Image Classification Dataset Distillation Image Classification

Open paper AskPDF arXiv

Dataset distillation (DD) compresses a large training set into a small synthetic set, reducing storage and training cost, and has shown strong results on general benchmarks. Decoupled DD further improves efficiency by splitting the pipeline into pretraining, sample distillation, and soft-label generation. However, existing decoupled methods largely rely on coarse class-label supervision and optimize samples within each class in a nearly identical manner. On fine-grained datasets, this often yields distilled samples that (i) retain large intra-class variation with subtle inter-class differences and (ii) become overly similar within the same class, limiting localized discriminative cues and hurting recognition. To solve the above-mentioned problems, we propose FD

^{2}

, a dedicated framework for Fine-grained Dataset Distillation. FD

^{2}

localizes discriminative regions and constructs fine-grained representations for distillation. During pretraining, counterfactual attention learning aggregates discriminative representations to update class prototypes. During distillation, a fine-grained characteristic constraint aligns each sample with its class prototype while repelling others, and a similarity constraint diversifies attention across same-class samples. Experiments on multiple fine-grained and general datasets show that FD

^{2}

integrates seamlessly with decoupled DD and improves performance in most settings, indicating strong transferability. Code is available at https://github.com/Guang000/FD2.

Abstract

Instance Segmentation

Explore similar work

FD2^22: A Dedicated Framework for Fine-Grained Dataset Distillation

Cross-Domain Dead Tree Detection via Knowledge Distillation in Aerial Imagery

SegmentAnyTreeV2: Scaling Transformer-Based Tree Instance Segmentation Across Sensors, Platforms, and Forests

FD $^2$ : A Dedicated Framework for Fine-Grained Dataset Distillation