Multi-Scale Center Proposal Network

• The paper introduces a family of architectures that generate scale-aware proposals using various center-based and location-based methods.
• It employs diverse mechanisms—such as pixel-wise regression, anchor-based sliding windows, and anchor-free corner pairing—to handle large scale variations.
• Empirical studies demonstrate enhanced proposal recall and improved localization quality across object detection and pathology MIL applications.

Multi-Scale Center Proposal Network, as an Editor’s term, denotes a family of proposal-generation and proposal-like spatial-guidance architectures that couple explicit scale handling with spatially localized candidate formation. In the literature considered here, the spatial primitive varies substantially: SPOP-net predicts a bounding box from each pixel; MS-CNN and the position-sensitive proposal network attach proposal predictors to feature-map locations and sliding windows; FastMask predicts object masks from centered sliding windows over a learned feature pyramid; CPN forms anchor-free proposals by pairing class-consistent corners; and MSPN generates coarse spatial guidance maps over remapped whole-slide-image grids rather than explicit boxes. The unifying design problem is stable proposal quality under large scale variation, but the term does not identify a single canonical architecture or a single center representation (Jie et al., 2016, Cai et al., 2016, Lu et al., 2018, Hu et al., 2016, Duan et al., 2020, Wu et al., 2 Feb 2026).

1. Scope and taxonomic boundaries

In a narrow modern sense, a center-based proposal network predicts explicit object centers, center heatmaps, or center-ness scores. Several systems associated with the broader multi-scale proposal literature are instead location-centered, support-aligned, or center-constrained rather than explicitly center-supervised. MS-CNN places anchors centered at sliding-window locations on multiple backbone layers; the position-sensitive proposal network derives anchors directly from sliding windows on decoded pyramid levels; FastMask treats each sliding window as an object-centered candidate if the object center lies in the central $10 \times 10$ region; SPOP-net predicts a proposal from every foreground pixel; CPN uses corner pairs rather than centers; and MSPN is explicitly described as a close conceptual match rather than a direct match, because it learns coarse guidance maps rather than proposal boxes or center coordinates (Cai et al., 2016, Lu et al., 2018, Hu et al., 2016, Jie et al., 2016, Duan et al., 2020, Wu et al., 2 Feb 2026).

System	Spatial unit	Scale mechanism
SPOP-net	Foreground pixel to box	Large/small localization networks plus original and $2\times$ test scales
MS-CNN	Feature-map location with centered anchor	det-8, det-16, det-32, det-64 branches
Position-sensitive proposal network	Sliding window and support-aligned anchor	Decoder levels $D_2$ to $D_6$ plus window-to-anchor mapping
FastMask	Center-constrained sliding window	Recursive residual neck and two-stream densification
CPN	Top-left and bottom-right corner pair	Geometry-flexible anchor-free pairing plus multi-scale testing
MSPN	Coarse grid cell guidance	Stacked CGNs over user-defined FOVs

This taxonomy matters because “multi-scale center proposal network” can otherwise be misconstrued as a single architectural template. The papers instead span dense box regression, anchor-based proposal heads, one-shot segment proposal, anchor-free corner pairing, and coarse-to-fine weakly supervised guidance. The shared theme is multi-scale spatial selection; the operational realization differs by task and supervision regime (Jie et al., 2016, Cai et al., 2016, Lu et al., 2018, Hu et al., 2016, Duan et al., 2020, Wu et al., 2 Feb 2026).

2. Spatial proposal primitives and center-like parameterizations

SPOP-net is a dense pixel-wise proposal model. For each pixel $i$, it predicts a 4D vector of normalized box coordinates corresponding to the object containing that pixel, and it is trained so that the network effectively predicts offsets relative to the pixel itself rather than absolute image-position-dependent boxes. This gives every foreground pixel an opportunity to vote for the enclosing object, which is why the method is often described as center-relative in spirit even though it does not produce a center heatmap (Jie et al., 2016).

MS-CNN uses an anchor-based sliding-window formulation, but its anchors are explicitly centered at the sliding window on layer $m$. Proposal heads are attached to multiple output layers, so proposal generation is tied to the spatial support of each branch. The position-sensitive proposal network preserves a similar spatial attachment, but makes anchor geometry support-aware: for a window with width $w$, height $h$, and feature-map stride $s$, the image-space anchor size is $(w \cdot s) \times (h \cdot s)$. Its position-sensitive score maps then partition each proposal support into $2\times$0 bins, preserving translation-variance for localization and translation-invariance for objectness (Cai et al., 2016, Lu et al., 2018).

FastMask departs from box-first proposal generation. On a feature map downscaled by factor $2\times$1, a sliding window of size $2\times$2 corresponds to a patch of $2\times$3 in the original image. A window is treated as a positive sample only if it fully contains the object, the object scale lies in $2\times$4 relative to the window, and the object center lies in the central $2\times$5 region. The head predicts confidence, a segmentation mask, and an attention map, so the proposal is mask-valued rather than box-valued (Hu et al., 2016).

CPN uses a different anchor-free primitive: the proposal is the Cartesian pairing of a top-left and a bottom-right corner of the same class, subject to $2\times$6 and $2\times$7. This is not center-based, but it remains location-based and highly geometry-flexible. MSPN is farther from proposal generation in the detection sense: it remaps high-magnification WSI features into coarse grids, predicts a one-channel coarse guidance map over those cells, and propagates the guidance back to the original patches as a patch-wise attention map. The paper explicitly states that it never predicts a list of centers, never outputs boxes, never performs NMS, and is not trained with localization supervision (Duan et al., 2020, Wu et al., 2 Feb 2026).

3. Multi-scale design strategies

The most direct scale-specialized design in the surveyed literature is SPOP-net’s divide-and-conquer formulation. It splits objects by area into large objects with area $2\times$8 pixels and small objects with area $2\times$9 pixels, trains separate localization networks for the two regimes, and fuses their outputs using a large-/small-size weighting network with per-pixel confidence $D_2$0. Test-time inference is run at the original image scale and at an image enlarged by factor $D_2$1, which further improves small-object coverage (Jie et al., 2016).

MS-CNN internalizes scale specialization into the backbone. Its proposal sub-network attaches heads to det-8, det-16, det-32, and det-64, and only a subset $D_2$2 of the training samples contributes to the loss of detection layer $D_2$3. Lower branches therefore specialize to small objects and higher branches to large objects, with receptive fields that more closely match object extent than a single deep proposal head (Cai et al., 2016).

The position-sensitive proposal network combines a ResNet-50 encoder with a top-down decoder producing $D_2$4 and $D_2$5 at strides $D_2$6. Scale handling is distributed across multiple design choices: the decoded feature pyramid, level-specific sliding-window profiles, direct window-to-anchor mapping, and a $D_2$7 large-kernel context module in the proposal head. The additional $D_2$8 level is introduced specifically to handle extremely large objects, while $D_2$9 carries very dense anchors for very small objects (Lu et al., 2018).

FastMask replaces image-pyramid evaluation with a recursive, weight-shared residual neck that generates a feature pyramid from one semantic body feature map. To densify scale coverage, it uses a two-stream construction that produces features downscaled by 16 and 24, then applies shared top convolutions, the shared neck, and the shared head. CPN handles scale differently: it relies on the intrinsic scale flexibility of corner-defined boxes, the multi-resolution fusion of stacked Hourglass backbones, and multi-scale test-time evaluation at $D_6$0 and $D_6$1 (Hu et al., 2016, Duan et al., 2020).

MSPN translates the same problem into WSI MIL. Rather than extracting multiple scanner-defined magnifications, it selects user-defined FOVs, remaps high-magnification features into progressively coarser pseudo-grids, and stacks coarse guidance networks from larger to smaller FOVs. In the reported implementation, three FOVs—$D_6$2—are used for fair comparison with common $D_6$3 multi-scale MIL, while an ablation studies one to five FOVs up to $D_6$4 (Wu et al., 2 Feb 2026).

4. Supervision, loss design, and proposal selection

SPOP-net uses foreground-filtered box regression, so only foreground pixels contribute to localization loss. It further defines large-only and small-only localization losses, and a confidence network jointly predicts foreground/background objectness and the probability that a foreground pixel belongs to a large object. Proposal ranking is then based on objectness, after optional superpixel-based shrinkage and expansion refinements, followed by NMS with threshold $D_6$5 (Jie et al., 2016).

MS-CNN uses a multi-branch proposal loss in which each branch contributes classification and smoothed $D_6$6 localization terms, but only for scale-assigned samples. An anchor is positive if its IoU with a ground-truth box is at least $D_6$7, preliminary negative if IoU $D_6$8, and anchors in the intermediate range are discarded. Hard negative mining is implemented through bootstrapping or mixture sampling, and the combined outputs of all proposal branches are declared as the final proposal detections (Cai et al., 2016).

The position-sensitive proposal network uses cross-entropy for objectness and smooth $D_6$9 for regression on sampled anchors from all pyramid levels. An anchor is positive if IoU with any ground-truth box is $i$0, negative if IoU $i$1, and the highest-IoU anchor for each ground-truth box is additionally assigned positive only if its IoU exceeds $i$2. Per image, anchors are sampled with a 1:1 positive/negative balance (Lu et al., 2018).

FastMask supervises confidence on all windows, but segmentation and attention only on positive windows. Its positive definition is based on containment, relative scale $i$3, and centeredness, not anchor IoU. CPN combines four losses—corner detection, corner offset, binary proposal classification, and multi-class classification. For the binary proposal stage and the class stage, positives are determined by IoU threshold $i$4; at inference, proposals with binary score above $i$5 are retained for class prediction, soft-NMS is applied, and the top 100 detections are kept (Hu et al., 2016, Duan et al., 2020).

MSPN is supervised at the downstream slide level. It is inserted before the MIL aggregator, the foundation-model feature encoder remains frozen, and classification uses cross-entropy. For prognosis prediction, the appendix provides a hazard-survival formulation with censored, uncensored, and combined survival losses. This makes MSPN structurally distinct from detection proposal networks: it is optimized end-to-end without localization supervision, and its coarse spatial maps serve as dense regional gating rather than explicit proposal outputs (Wu et al., 2 Feb 2026).

5. Representative architectures and empirical characteristics

SPOP-net emphasizes localization quality at high IoU. On PASCAL VOC 2007, it is reported as consistently among the best and usually the best at IoU $i$6, ranks first in average recall over IoU range $i$7 to $i$8, and is strongest in ABO especially when proposal count exceeds 500. When used with Fast R-CNN on VOC 2007 test, it yields $i$9. Its reported runtime is $m$0 s/image, with $m$1 s for one network at original scale and $m$2 s for one network at enlarged scale (Jie et al., 2016).

MS-CNN emphasizes proposal recall under strict proposal budgets. On KITTI validation, proposal generation reaches about $m$3 recall with only 100 proposals. Runtime is reported as 10 fps on KITTI images $m$4 and 15 fps on Caltech images $m$5, with the submitted KITTI model around $m$6 s/image. Its scale-specialized branches are empirically complementary: det-8 is best for the smallest height range, det-16 for the next, det-32 for medium-to-large, and det-64 for the largest (Cai et al., 2016).

The position-sensitive proposal network emphasizes average recall and localization-sensitive proposal scoring. On PASCAL VOC, baseline $m$7 increases to $m$8 for GCN-NS with position sensitivity; on COCO, baseline $m$9 increases to $w$0. The headline claim is a $w$1 improvement at 1,000 proposals on PASCAL VOC and $w$2 on COCO, with inference time $w$3 ms for input image size $w$4 (Lu et al., 2018).

FastMask shows how one-shot multi-scale segment proposal can be made efficient. On COCO validation for segmentation proposals, FastMask with two streams and Res39 reports AR@100 $w$5 and AR@1k $w$6. The “FastMask-acc” configuration runs at $w$7 s/image, while “FastMask-fast” runs at $w$8 s/image; the latter is described as near real time, about 13 FPS on $w$9 images. Its ablations attribute major gains to the residual neck and attentional head, with AR@100 rising from 24.8 to 29.3 when the attentional head replaces the standard head (Hu et al., 2016).

CPN couples high-recall anchor-free proposal generation with explicit regional verification. On COCO test-dev, CPN with HG-104 reports 47.0 AP in single-scale testing and 49.2 AP in multi-scale testing. A lighter DLA-34 variant achieves 41.6 AP at 26.2 FPS with flip and 39.7 AP at 43.3 FPS without flip. Proposal recall analysis reports overall AR $h$0, $h$1 for very large objects, and strong recall for extreme aspect ratios. Its average false discovery is reported as 30.6 for HG-104, compared with 37.8 for CornerNet and 32.4 for CenterNet (Duan et al., 2020).

MSPN transfers the multi-scale proposal idea into weakly supervised pathology. For ABMIL + CONCH, the single-scale baseline reports ER/PR/HER2/CRC-Surv $h$2, whereas ABMIL + MSPN reaches $h$3. The paper reports average MSPN improvements over corresponding non-MSPN variants of about 1.84% for ABMIL, 2.21% for DSMIL, 3.53% for CLAM-SB, and 2.41% for CLAM-MB. In compute terms, ABMIL + MSPN uses 2.18M parameters and 17.71G FLOPs, compared with 2.61M and 38.37G for ABMIL with concatenation or CSA, 2.93M and 84.01G for TransMIL, and 85.83M and 503.30G for HAG-MIL (Wu et al., 2 Feb 2026).

6. Conceptual limits, common misconceptions, and broader significance

A persistent misconception is that any multi-scale proposal architecture is necessarily an explicit center-based detector. The surveyed systems show otherwise. SPOP-net is dense pixel-to-box regression, not center-heatmap prediction. MS-CNN and the position-sensitive proposal network are anchor-based and centered on feature-map locations, not anchor-free center-point models. FastMask is a center-constrained mask proposer without anchor boxes or explicit center heatmaps. CPN is corner-based rather than center-based. MSPN is not a proposal network in the object-detection sense at all; it is a progressive coarse-guidance module for WSI MIL (Jie et al., 2016, Cai et al., 2016, Lu et al., 2018, Hu et al., 2016, Duan et al., 2020, Wu et al., 2 Feb 2026).

A second misconception is that multi-scale proposal quality must be obtained from an image pyramid. FastMask explicitly replaces multi-shot image-pyramid inference with a learned feature pyramid, and MSPN derives coarser context directly from the highest-magnification WSI features. MS-CNN and the position-sensitive proposal network internalize scale through multi-level proposal heads and decoded feature pyramids. By contrast, SPOP-net and CPN still use explicit test-time multi-scale evaluation, showing that the literature supports both internal and external scale mechanisms (Hu et al., 2016, Wu et al., 2 Feb 2026, Cai et al., 2016, Lu et al., 2018, Jie et al., 2016, Duan et al., 2020).

The papers collectively suggest that the most reusable principles are not tied to a single choice of center representation. Scale-specialized heads, support-aligned proposal geometry, dense over-complete candidate generation, large-context modeling, and a later selection or verification stage recur across otherwise different formulations. In detection, these principles appear as branch-specific anchors, position-sensitive bins, corner-pair proposal sets, and proposal classifiers. In pathology MIL, they appear as dense coarse guidance maps that modulate fine-scale patch representations before bag-level inference. A plausible implication is that “multi-scale center proposal network” is best treated as a design family defined by scale-aware spatial candidate formation and refinement, rather than by any one canonical output such as centers, anchors, corners, masks, or coarse cells (Cai et al., 2016, Lu et al., 2018, Duan et al., 2020, Wu et al., 2 Feb 2026).