TimberSeg 1.0: Log Segmentation in Forestry

• TimberSeg 1.0 is a densely annotated RGB dataset designed for segmenting accessible logs in cluttered and realistic forestry scenes.
• The dataset comprises 220 images with 2500 individually segmented logs, providing both bounding boxes and pixel-level masks to facilitate autonomous log grasping.
• Benchmark analyses reveal that transformer-based methods like Mask2Former outperform axis-aligned and rotation-aware models, highlighting the advantages of pixel-level reasoning.

TimberSeg 1.0 most commonly denotes the densely annotated RGB dataset introduced for instance segmentation of graspable logs in realistic forestry operations, together with the benchmark established on that dataset for autonomous log grasping (Fortin et al., 2022). It was created to address the perception bottleneck in forwarders and wood loaders, where logs are elongated, randomly oriented, cluttered, partially occluded, and observed under adverse outdoor conditions. In its original formulation, TimberSeg 1.0 comprises 220 images with 2500 individually segmented logs and provides both bounding boxes and pixel-level instance masks for the single class “log” (Fortin et al., 2022). Later literature and secondary syntheses reuse the same label for several timber- and tree-segmentation pipelines, so the term has acquired a context-dependent meaning beyond its initial dataset definition (Zolotarev et al., 18 Mar 2025).

1. Original definition and problem setting

TimberSeg 1.0 was introduced to support perception for autonomous or operator-assist log picking, with the explicit focus on detecting and segmenting only the logs that are on top of piles and accessible to a forwarder or loader (Fortin et al., 2022). This scope is narrower than generic log detection: buried or inaccessible logs are intentionally left unannotated. The dataset spans viewpoints and environments representative of real operations, including forest scenes with scattered logs in freshly cut areas and roadside scenes with neatly piled logs ready for transport.

Its composition reflects the geometric and visual properties that make the task difficult. The dataset contains 220 RGB images with 2500 individually segmented log instances; the number of instances per image ranges from 1 to 29, with an average of 11.3 logs per image. Logs are strongly elongated: the aspect ratio distribution resembles a log-normal with mean 8.3 and a tail up to 24, far from typical 2:1 or 1:2 anchor ratios. Orientations are broadly distributed with a peak around 90 degrees and otherwise no dominant orientation. Scenes exhibit clutter and overlap, with frequent partial occlusion by other logs or branches, and they include snow, sun glare, and night operation (Fortin et al., 2022).

These properties are central to the benchmark’s significance. The dataset was explicitly designed to test detectors and segmenters that must operate under elongated geometry, arbitrary orientation, heavy overlap, and outdoor domain shift. The original study therefore framed TimberSeg 1.0 not merely as a corpus of forestry imagery but as a stress test for instance segmentation methods whose inductive biases had largely been developed on less anisotropic object classes.

2. Data sources, annotation, and protocol

The underlying imagery was assembled from three types of sources. First, three VIOFO A129 Pro Duo 4K dashcams were installed in forwarders operating near Lake Saint-Jean, Quebec, Canada; each dashcam had two cameras mounted inside the cabin, one forward-facing and one sideways-facing, and recorded hundreds of hours over multiple months and weather conditions at 4K, 2K, and 1080p. Second, a Canon EOS M50 camera was used to capture still images in lumberyards at sawmills and paper mills, including top-down views of loaded timber truck trailers. Third, complementary images were taken from publicly available online videos to increase geographic diversity (Fortin et al., 2022).

TimberSeg 1.0 is densely annotated with both bounding boxes and pixel-level instance masks. Masks were created as polygon segmentations using SuperAnnotate. A key annotation decision was to label only the topmost, accessible logs. To accelerate labeling, Mask2Former trained on earlier data was used to pre-annotate masks, and human annotators corrected these suggestions. The paper does not report inter-annotator agreement or formal QA procedures. It also does not state that rotated bounding boxes are part of the ground truth; bounding boxes are included, but their parameterization is not explicitly described (Fortin et al., 2022).

For model training, the images underwent large-scale jittering augmentation: random resize in $[0.1, 2.0]$ scale, followed by cropping to $1024 \times 1024$ pixels with gray padding, and horizontal flipping. Images were normalized using ImageNet RGB mean and standard deviation. Initial training details mention training for 8000 iterations with a test set comprising 20% of images, whereas the final reported results use 5-fold cross-validation, with metrics reported as mean $\pm$ standard deviation across folds (Fortin et al., 2022).

The dataset and code are available at https://github.com/norlab-ulaval/logpiles_segmentation. The source paper states that the dataset will be made public upon publication, but specific licensing terms, DOI, and usage restrictions are not specified (Fortin et al., 2022).

3. Benchmark methodology and model families

Evaluation on TimberSeg 1.0 uses COCO-style mask metrics for single-class “log” detection. The reported quantities are mean Average Precision on masks across IoU thresholds, $\operatorname{AP50}$, recall, $F1$-score, and Precision–Recall curves for mask IoU thresholds. The formulas given in the source are:

$$\operatorname{IoU}(A,B)=\frac{|A\cap B|}{|A\cup B|}$$

$\operatorname{AP}=\int_0^1 p(r)\,dr$

$\operatorname{mAP}=\frac{1}{N}\sum_{i=1}^N AP_i$

$$F1=\frac{2\cdot(\text{precision}\cdot \text{recall})}{\text{precision}+\text{recall}}$$

Predictions are filtered to the 100 highest-scoring detections at evaluation time, and mask IoU is the basis for AP, PR curves, and thresholded measures. The benchmark does not define or use rotated IoU for bounding boxes; evaluation focuses on instance masks rather than oriented boxes (Fortin et al., 2022).

Three baseline families were compared. The first is an axis-aligned region-based method, Mask R-CNN with a ResNeXt-101-FPN backbone. The second is a rotation-aware region-based method, Rotated Mask R-CNN, which uses Rotation Region Proposal Network and RRoIAlign; its anchors were configured with 30-degree orientation intervals and aspect ratios of 4, 8, 12, 16, and box regression uses parameters $(x, y, w, h, \theta)$. The third is an attention-based transformer method, Mask2Former with a Swin Transformer-B backbone, pretrained on ImageNet-1k or ImageNet-22k and COCO (Fortin et al., 2022).

The benchmark is methodologically interesting because each architecture expresses a different stance on orientation handling. Axis-aligned proposals rely on rectangular ROIs from predefined anchors, which is problematic when elongated logs at arbitrary angles cause ROIs to include multiple objects and background. Rotation-aware proposals reduce this mismatch by explicitly modeling angle. Mask2Former removes region proposals altogether and instead uses multi-scale deformable attention over the full feature maps and masked attention in the decoder; in the TimberSeg study this was presented as a way to reason directly at the pixel level with global context, without a rectangular ROI inductive bias (Fortin et al., 2022).

All baselines were trained in Detectron2 on PyTorch using a single machine with an NVIDIA RTX 3090 GPU, Intel Core i9-10900KF CPU, and 64 GB RAM. Mask R-CNN used SGD with momentum 0.9, weight decay 0.0001, initial learning rate 0.001, batch size 8, and 8000 iterations. Rotated Mask R-CNN retained the same backbone and optimization scheme while replacing RPN/ROIAlign with RRPN/RRoIAlign. Mask2Former used AdamW with weight decay 0.02, a 9-layer decoder across pyramid resolutions, and 100 object queries (Fortin et al., 2022).

4. Reported performance on TimberSeg 1.0

The empirical comparison shows a large separation between axis-aligned, rotation-aware, and transformer-based approaches. The headline result is that Mask2Former with Swin-B pretrained on ImageNet-22k achieves $1024 \times 1024$0, substantially above the region-based baselines (Fortin et al., 2022).

Model	Mask metrics	fps
Mask R-CNN (X101-FPN)	mAP 19.03 ± 3.21; AP50 36.10 ± 4.45; Recall 28.60 ± 3.98; F1 0.23 ± 0.04	12.55
Rotated Mask R-CNN (X101-FPN, RRPN + RRoIAlign)	mAP 31.83 ± 3.26; AP50 52.78 ± 4.33; Recall 36.95 ± 3.14; F1 0.34 ± 0.03	5.66
Mask2Former (Swin-B, ImageNet-1k)	mAP 56.05 ± 3.12; AP50 82.97 ± 1.99; Recall 64.06 ± 3.05; F1 0.60 ± 0.03	8.47
Mask2Former (Swin-B, ImageNet-22k)	mAP 57.53 ± 3.37; AP50 84.28 ± 2.44; Recall 65.16 ± 3.40; F1 0.61 ± 0.03	8.47

The study interprets these numbers as evidence that axis-aligned proposals underperform because they capture large swathes of irrelevant pixels and suffer under non-maximum suppression in dense piles. Rotation-aware proposals tighten the ROI around elongated logs and improve accuracy by roughly 12–13 mAP points over axis-aligned Mask R-CNN, but they remain constrained by the single-rectangle abstraction and by the computational cost of a larger anchor set. Mask2Former performs best because it reasons at the pixel level with global context and is therefore better suited to elongated, touching instances whose contours are interwoven (Fortin et al., 2022).

The qualitative analysis is consistent with the quantitative ranking. Mask2Former produced high-quality masks on clean roadside piles and forest scenes, and was described as robust under sun glare, snow-covered scenes, and night-time operation with vignetting from artificial lights. The source paper also notes that Mask2Former’s PR curves remain strong up to mask IoU $1024 \times 1024$1; at IoU $1024 \times 1024$2, precision above 80% can be maintained at recall around 80%. Using IoU $1024 \times 1024$3 as a grasp success proxy yields a precision of 84.39% for the best Swin-B 22k model, with the argument that the grapple jaw is several times wider than the log diameter (Fortin et al., 2022).

The study further reports practical trade-offs. Training time per fold was approximately 45 minutes for Mask2Former, compared with approximately 2.5 hours for Mask R-CNN and approximately 5 hours for Rotated Mask R-CNN. Data scaling with Mask2Former followed a power law: with a fixed 44-image test set, doubling the number of training images yielded approximately 6.87% absolute mAP improvement on average until plateau. Resolution scaling exposed an accuracy–latency frontier: reducing input size to $1024 \times 1024$4 increased speed to 24.05 fps at the cost of precision, while $1024 \times 1024$5 delivered an approximately 3× speedup versus $1024 \times 1024$6 with only an approximately 5% precision drop (Fortin et al., 2022).

5. Operational implications, limitations, and resources

Within the original use case, TimberSeg 1.0 is not only a benchmark dataset but also a perception substrate for downstream grasp planning. The paper argues that instance masks provide precise centers of mass and principal axes via mask moments, enabling grasp pose estimation by computing the mask centroid, principal axis, and width estimates, then sampling grasp candidates around the centroid aligned normal to the principal axis or with heuristic spacing along the axis. In a broader perception–planning–control pipeline, the segmenter identifies accessible instances, grasp candidates are evaluated against reachability and collision constraints, and grasp success can be improved via closed-loop servoing using real-time segmentation (Fortin et al., 2022).

The limitations are equally explicit. TimberSeg 1.0 is modest in size at 220 images and 2500 instances, even though it is diverse across environments and conditions. Only top, accessible logs are annotated; extreme occlusion or partially buried logs remain unannotated by design. Formal QA and inter-annotator agreement are not reported. Evaluation is single-frame, without temporal modeling or tracking. Although qualitative robustness to glare, snow, and night conditions is shown, systematic sensitivity analyses to lighting and weather are not provided, and out-of-distribution behavior beyond the collected domains is not assessed (Fortin et al., 2022).

The future directions proposed around TimberSeg 1.0 follow directly from these constraints. They include larger datasets and continued labeling; semi- and self-supervised pretraining on the large unlabeled dashcam video corpus, including Masked Autoencoders; advanced augmentation such as copy-paste; temporal modeling through explicit tracking, Bayesian filtering, or direct video instance segmentation; rotation-aware transformers or hybrid methods; and multi-modal sensing such as RGB-D or LiDAR for improved robustness in extreme conditions and for 3D grasp planning (Fortin et al., 2022).

6. Terminological reuse and relation to adjacent forestry vision

Although TimberSeg 1.0 originated as the Fortin et al. dataset for RGB log instance segmentation, later literature and derivative syntheses reuse the label in broader or different senses. One direct extension appears in “Deep Unsupervised Segmentation of Log Point Clouds,” where “TimberSeg 1.0” denotes an unsupervised Point Transformer-based method for segmenting the surface of a single debarked log from raw laser-scanned point clouds in sawmill settings. In that usage, the model outputs per-point weights and centreline vectors and reports whole-test-set precision 98.88%, recall 98.45%, and IoU 97.35% (Zolotarev et al., 18 Mar 2025). A different reuse appears in “A Low-Cost Machine Learning Approach for Timber Diameter Estimation,” which fine-tunes YOLOv5 on public TimberSeg 1.0 imagery for single-class log detection and coarse diameter binning from bounding-box width; on a held-out test set of 21 images with 208 labeled instances it reports precision 0.656, recall 0.577, [email protected] 0.640, and [email protected]:0.95 0.356 (Fard et al., 23 Jul 2025).

Some secondary syntheses use the name even more loosely. A blueprint derived from aerial image time series for tree species semantic segmentation is presented under the label “TimberSeg 1.0,” with best reported mIoU $1024 \times 1024$7 for Processor+U-Net with ResNet-101 and hierarchical loss (Ramesh et al., 2024). Another synthesis explicitly states that the name is not used in the source paper and applies it editorially to the best pipeline for wood-rot semantic segmentation on spruce log crosscuts, where InternImage-H-UPerNet on the augmented dataset reaches IoU (All) 0.71 and F1 (All) 0.85 (Kammerbauer et al., 2024). This suggests that “TimberSeg 1.0” now functions both as a specific dataset name and as a portable label for timber-oriented segmentation systems.

The broader research context is correspondingly heterogeneous. Adjacent forestry-vision work includes multi-class graph cut crown delineation from LiDAR point clouds (Williams et al., 2019), Faster R-CNN plus 3D-FCN pipelines for airborne LiDAR tree detection and crown segmentation (Windrim et al., 2018), weakly supervised tree instance segmentation in ALS point clouds (Destouches et al., 21 Aug 2025), scalable distributed crown segmentation for forest-scale small-footprint airborne LiDAR (Hamraz et al., 2017), non-parametric deciduous-tree segmentation from ALS (Hamraz et al., 2017), and wood-versus-leaf TLS segmentation across diverse European forests (Owen et al., 6 Mar 2025). Within that landscape, the original TimberSeg 1.0 remains distinctive for centering log instance masks in cluttered forestry scenes and for demonstrating, on that task, the failure mode of rectangular region proposals and the relative advantage of attention-based pixel-level reasoning (Fortin et al., 2022).