Mask Contour Reformer in HoliTracer

• Mask Contour Reformer (MCR) is a rule-based module that reforms raw segmentation contours into regularized polygon sequences for improved vector refinement.
• It employs a simplify-then-reconstruct approach using the Douglas-Peucker algorithm and uniform interpolation to align reconstructed polygons with ground truth.
• MCR operates differently in training and prediction, generating detailed training samples while bridging raster segmentation and learned polygon refinement.

Mask Contour Reformer (MCR) is, in "HoliTracer: Holistic Vectorization of Geographic Objects from Large-Size Remote Sensing Imagery" (Wang et al., 22 Jul 2025), a polygon reconstruction and training-sample generation module positioned between semantic segmentation and learned polygon refinement. Its function is to convert raw segmentation-mask boundaries into a regularized polygon point sequence suitable for the Polygon Sequence Tracer (PST), because contours extracted directly from segmentation are described as "dense and irregular" and make "direct refinement and correspondence with ground truth polygons challenging." In this usage, MCR is not the final vectorizer by itself and is not a learned network; it is a rule-based contour reforming and alignment procedure that bridges raster segmentation output and trainable polygon refinement.

1. Definition and role within HoliTracer

HoliTracer is organized into three components: Context Attention Net (CAN) for segmentation on large-size remote sensing imagery (RSI), Mask Contour Reformer (MCR) for polygon contour reconstruction, and Polygon Sequence Tracer (PST) for polygon refinement and vertex identification (Wang et al., 22 Jul 2025). The operational pipeline is therefore:

Large-size image $\rightarrow$ CAN segmentation mask $\rightarrow$ MCR reconstructed polygon sequence $\rightarrow$ PST refined polygon + detected vertices $\rightarrow$ final vector polygon

Within that pipeline, MCR is introduced "to generate more robust and trainable polygon inputs." It "reconstructs polygons and aligns them with ground truth to produce training samples," and in the introduction it "reconstructs irregular polygon contours while ensuring alignment with ground truth polygons through a bidirectional matching mechanism." The detailed mechanism described in the paper, however, is rule-based rather than learned: contour extraction, simplification, interpolation, nearest-point matching, and sequence reconstruction.

A frequent misconception is to treat MCR as the vectorization system itself. The paper states the opposite structure. MCR solves the representation and supervision-alignment problem, whereas PST performs learned geometric refinement. Another misconception is to interpret MCR as a CNN or graph module. Factually, nothing in the paper suggests that MCR is neural; it operates on geometry extracted from segmentation masks rather than on raw image features or CNN feature maps.

2. Inputs, outputs, and operational modes

MCR takes as its primary input the segmentation contour extracted from the predicted mask,

$S = \{s_1, s_2, \dots, s_n\},$

where each point is

$s_i = (x_i, y_i).$

During training it also uses the ground-truth polygon

$G = [g_1, g_2, \dots, g_M],$

together with the simplification tolerance $\epsilon$ and interpolation distance $l$ (Wang et al., 22 Jul 2025).

The module behaves differently in training and prediction.

Mode	Inputs	Outputs
Training	$S$, $\rightarrow$0, $\rightarrow$1, $\rightarrow$2	$\rightarrow$3, $\rightarrow$4, $\rightarrow$5
Prediction	$\rightarrow$6, $\rightarrow$7, $\rightarrow$8	$\rightarrow$9

During training, MCR outputs the reconstructed segmentation polygon

$\rightarrow$0

the reconstructed ground-truth polygon with pointwise correspondence $\rightarrow$1, and binary vertex labels

$\rightarrow$2

During prediction, the paper states: "In the prediction stage, MCR applies only the reconstruction process to segmentation results, outputting $\rightarrow$3 without matching or ground truth interpolation" (Wang et al., 22 Jul 2025).

This training/prediction asymmetry is central. MCR is simultaneously a polygon reconstructor and a supervision generator, but only the reconstruction branch survives at inference. A plausible implication is that the paper treats contour reforming and label construction as separable concerns: geometric regularization is required both at train time and test time, whereas pointwise correspondence to annotation is needed only for learning.

3. Geometric reconstruction procedure

MCR begins by extracting polygon contours from segmentation results using TC89-KCOS:

$\rightarrow$4

The paper then applies a "simplify-then-reconstruct approach" with the Douglas-Peucker (DP) algorithm using tolerance $\rightarrow$5, producing

$\rightarrow$6

with $\rightarrow$7 and $\rightarrow$8 (Wang et al., 22 Jul 2025).

After simplification, MCR reconstructs the polygon by interpolating points along each edge of $\rightarrow$9 at fixed distance $\rightarrow$0. The reconstructed polygon is

$\rightarrow$1

with

$\rightarrow$2

and

$\rightarrow$3

The supplementary description expands this edge-wise interpolation with

$\rightarrow$4

$\rightarrow$5

The paper notes a minor notation inconsistency in the supplementary pseudocode, where a symbol $\rightarrow$6 appears in line 7 without separate definition; the intended fixed spacing parameter is clearly $\rightarrow$7.

The significance of this stage is not merely contour simplification. DP alone yields a reduced set of dominant vertices, but PST requires a reasonably ordered, regularly spaced polygon sequence. MCR therefore regularizes sampling density after simplification. This converts a dense, pixel-driven contour into a uniformly sampled polygonal sequence suitable for sequence processing.

The method is explicitly restricted to geometric primitives already present in the segmentation contour. The paper states that MCR uses mask contours, TC89-KCOS contour extraction, Douglas-Peucker simplification, uniform interpolation along polygon edges, nearest-neighbor geometric matching to ground-truth vertices, sequence reconstruction, and binary vertex labeling. It does not use skeletons, dynamic programming inside MCR itself, graph neural networks, dynamic contour evolution, polygon fitting by optimization, or learned contour deformation.

4. Ground-truth alignment and supervision generation

The training-only part of MCR aligns the reconstructed segmentation polygon $\rightarrow$8 with the annotation polygon

$\rightarrow$9

For each ground-truth vertex $S = \{s_1, s_2, \dots, s_n\},$0, MCR finds the closest point in $S = \{s_1, s_2, \dots, s_n\},$1:

$S = \{s_1, s_2, \dots, s_n\},$2

with Euclidean distance

$S = \{s_1, s_2, \dots, s_n\},$3

The paper says that the unique matched points $S = \{s_1, s_2, \dots, s_n\},$4 are designated as vertices, with ordered indices

$S = \{s_1, s_2, \dots, s_n\},$5

where $S = \{s_1, s_2, \dots, s_n\},$6 if multiple ground-truth vertices map to the same reconstructed point (Wang et al., 22 Jul 2025).

Once these matched points are determined, MCR assigns binary vertex labels

$S = \{s_1, s_2, \dots, s_n\},$7

forming

$S = \{s_1, s_2, \dots, s_n\},$8

MCR then constructs a reconstructed ground-truth polygon sequence $S = \{s_1, s_2, \dots, s_n\},$9 with exactly the same number of points as $s_i = (x_i, y_i).$0. For each pair of consecutive matched vertices $s_i = (x_i, y_i).$1 and $s_i = (x_i, y_i).$2, the number of intermediate non-vertex points is

$s_i = (x_i, y_i).$3

with wraparound $s_i = (x_i, y_i).$4 for the closed polygon. The corresponding ground-truth edge $s_i = (x_i, y_i).$5 is interpolated using

$s_i = (x_i, y_i).$6

Concatenation gives

$s_i = (x_i, y_i).$7

with the size constraint

$s_i = (x_i, y_i).$8

The paper explicitly states: "This ensures a one-to-one correspondence with $s_i = (x_i, y_i).$9."

This one-to-one pairing is the key interface to PST. PST defines ground-truth offsets as

$G = [g_1, g_2, \dots, g_M],$0

Those offsets, together with the binary labels $G = [g_1, g_2, \dots, g_M],$1, provide supervision for offset regression and vertex classification. Factually, MCR itself has no learned parameters and no loss function; its importance is that it makes PST supervision well-posed.

The terminology around matching requires care. The paper introduces MCR as aligning reconstructed polygons with ground truth "through a bidirectional matching mechanism," but the detailed equations only explicitly define nearest-point matching from $G = [g_1, g_2, \dots, g_M],$2 to $G = [g_1, g_2, \dots, g_M],$3, followed by sequence reconstruction. A plausible implication is that "bidirectional" in the broader framework description is not accompanied by a distinct bidirectional equation in the MCR section.

5. Hyperparameters, empirical role, and limitations

The paper specifies the main MCR hyperparameters as follows: simplification tolerance

$G = [g_1, g_2, \dots, g_M],$4

and interpolation distance

$G = [g_1, g_2, \dots, g_M],$5

(Wang et al., 22 Jul 2025). The algorithm itself is unchanged across buildings, water bodies, and roads; the object-type-specific adjustment lies in the choice of $G = [g_1, g_2, \dots, g_M],$6.

The paper does not provide a dedicated MCR-only ablation table, but it offers indirect evidence that contour reforming is necessary for vector quality. On WHU-Building, the reported comparison is: Baseline with PoLiS 3.83, CIoU 18.47, AP 58.75; Baseline + DP with PoLiS 4.02, CIoU 60.32, AP 58.42; and Baseline + PST with PoLiS 3.63, CIoU 82.30, AP 61.07. On GLH-Water, CIoU improves from 26.74 to 55.08 with Baseline + DP and to 57.88 with Baseline + PST. On VHR-Road, CIoU improves from 1.43 to 5.41 with Baseline + DP and to 6.10 with Baseline + PST (Wang et al., 22 Jul 2025). The paper interprets these results as showing that raw contour extraction preserves detail but is poor as vector representation, simplification helps substantially, and learned refinement on structured polygon input helps most.

Because PST depends on MCR’s outputs, these results are indirect rather than isolated evidence for MCR. The paper is explicit on this point: there is no separate ablation varying TC89-KCOS alone versus full MCR without PST, no study of different $G = [g_1, g_2, \dots, g_M],$7, no study of different $G = [g_1, g_2, \dots, g_M],$8, and no comparison of different matching strategies.

Several assumptions and limitations are visible from the formulation. First, the pseudocode uses wraparound for closed polygons, so MCR assumes a closed polygon contour. Second, it assumes that matched reconstructed points correspond to ground-truth vertices "in order." Third, nearest-point assignment is assumed sufficient for alignment. The paper also explicitly notes that

$G = [g_1, g_2, \dots, g_M],$9

when multiple ground-truth vertices map to the same reconstructed point, which means matching can collapse distinct annotation vertices when $\epsilon$0 is too coarse. It further notes a tradeoff between contour detail and vector simplicity, remarks that road vectorization remains challenging due to complex network structure, and implies sensitivity to sampling density through the object-specific choice of $\epsilon$1.

Equally important are the omissions. The paper does not discuss self-intersections, hole polygons or inner rings, explicit denoising of tiny spurs, connected-component filtering, topological repair, overlap handling among neighboring instances, or handling fragmented masks inside MCR itself. It also reports Ours MCR infer time: 132.78 s/sample in the supplementary compute table (Wang et al., 22 Jul 2025).

6. Terminological ambiguity and relation to adjacent contour methods

The acronym MCR is not stable across arXiv literature, and the term "Mask Contour Reformer" should therefore be localized carefully. In HoliTracer, MCR specifically means Mask Contour Reformer, a rule-based geometric module for polygon reconstruction from segmentation masks (Wang et al., 22 Jul 2025). In "Mask Consistency Regularization in Object Removal," by contrast, MCR means Mask Consistency Regularization: a training strategy for diffusion-based object removal that perturbs mask geometry through dilation and reshape and enforces denoising consistency under those perturbations (Yuan et al., 12 Sep 2025). In "MCR-Bionic Hand: Anatomical Structural Priors for Dexterous Manipulation," MCR refers to a dexterous-manipulation framework built around structural prior generation and muscle-mediated modulation, not to mask processing at all (Yang et al., 11 Jun 2026).

This naming ambiguity matters because several nearby research directions address contour reforming without using the same acronym. "SCR: Smooth Contour Regression with Geometric Priors" models contours as complex periodic functions and regresses a compact set of complex Fourier coefficients (Bahl et al., 2022). "ContourRend: A Segmentation Method for Improving Contours by Rendering" refines segmentation masks by sampling and classifying pixels around a predicted contour (Chen et al., 2020). "Contour Loss for Instance Segmentation via k-step Distance Transformation Image" adds contour-specific supervision to mask predictors through a differentiable truncated distance-transform construction (Guo et al., 2021). "Joint Object Contour Points and Semantics for Instance Segmentation" uses contour point detection as an auxiliary task in Mask R-CNN (Zhang et al., 2020). "Contour Refinement using Discrete Diffusion in Low Data Regime" proposes a mask-conditioned post-hoc contour refiner based on simplified discrete diffusion (Guan et al., 5 Feb 2026).

A useful distinction follows from these comparisons. HoliTracer’s MCR reforms contours at the interface between raster segmentation and polygon-sequence learning. It does not directly optimize contour smoothness in a spectral basis, render boundary pixels at high resolution, define a contour-aware loss for dense masks, or denoise a contour representation through diffusion. Its specific contribution is to convert irregular raster contours into a trainable polygon sequence and aligned supervision format for learned vector refinement. In that narrower and more exact sense, Mask Contour Reformer is best understood as a deterministic contour simplification, resampling, and alignment module embedded in a large-size RSI vectorization pipeline.