ParticleSAM: Dense Segmentation for Small Objects

• The paper introduces ParticleSAM, a near zero-shot adaptation of SAM for dense segmentation of small particles in recycling material images.
• It refines the inference pipeline with targeted mask filtering, enhanced feature resolution, and tuned decoder parameters to handle extreme small-object regimes.
• ParticleSAM leverages an automated synthetic dataset generation pipeline, yielding significant quantitative gains over the original SAM in mIoU and mAP metrics.

Searching arXiv for the ParticleSAM paper and its immediate methodological context. Searching for the original Segment Anything Model paper used as the base model. ParticleSAM is an adaptation of the Segment Anything Model for images containing small and dense objects, motivated by material quality monitoring in recycling processes for construction aggregates. The construction industry is identified as a major sector in terms of resource consumption, while recycled construction material has high reuse potential and aggregate quality monitoring is typically still performed with manual methods. Against that background, ParticleSAM targets images with hundreds of small particles, a regime for which existing segmentation methods are described as not directly applicable by design. The work couples a near zero-shot adaptation of SAM with a new dense multi-particle dataset simulated from isolated particle images through an automated data generation and labeling pipeline, and reports advantages over the original SAM in both quantitative and qualitative experiments (Zhou et al., 5 Aug 2025).

1. Problem formulation and domain context

ParticleSAM is positioned at the intersection of foundation-model segmentation and industrial visual inspection. Its immediate application is visual material quality control automation for recycled construction aggregates, where the relevant images contain many small particles, frequent contact between adjacent objects, and substantial occlusion. The paper frames this as a setting in which standard segmentation workflows are not well aligned with the image statistics: rather than a small number of prominent objects, the target images contain hundreds of fine-scale instances (Zhou et al., 5 Aug 2025).

The method is also presented as broader than a single industrial task. The same small-particle regime arises in domains with densely packed objects, controlled size distributions, and operational constraints on throughput. This suggests that the central problem is not only aggregate monitoring but segmentation under extreme object-count density, small object scale, and limited tolerance for latency.

2. Adaptation of the Segment Anything Model

ParticleSAM starts from the publicly released Segment Anything Model with a ViT-B backbone (Kirillov et al., 2023) and modifies the inference pipeline in three ways: mask filtering, feature extraction enhancement, and decoder/post-processing tuning (Zhou et al., 5 Aug 2025).

The first change concerns mask filtering. In standard SAM’s Automatic Mask Generator, small masks below a minimum area are dropped. ParticleSAM inverts this logic and removes the very largest masks, which are described as background “bleed-through” from the conveyor, while retaining masks only within a prescribed area band,

$A_{\min} \le A(\hat M_i) \le A_{\max}.$

The paper characterizes this as a “maximum-area filter” that removes spurious huge regions and speeds up inference on fine particles (Zhou et al., 5 Aug 2025).

The second change increases prompt-related feature resolution. SAM’s original image encoder samples features on a $32\times32$ grid for prompts; ParticleSAM increases this to a $64\times64$ grid, halving the sampling stride so that each token covers a smaller spatial patch. The stated purpose is to enrich local feature representation for small objects. The authors also experimented with splitting a $4096\times4096$ input into four quadrants, encoding them separately, and fusing decoder features. Although this improved very heavy-occlusion cases, it slowed inference by $\times5$, so the retained configuration uses a full-image pass with the finer $64\times64$ grid (Zhou et al., 5 Aug 2025).

The third change tunes decoder-side suppression and filtering. ParticleSAM adjusts the Non-Maximum Suppression IoU threshold $t_{\rm nms}$ and the stability threshold $t_{\rm stab}$ for mask duplicates. The final mask set $\{\hat M_i\}$ is filtered so that

$$\mathrm{IoU}(\hat M_i,\hat M_j) < t_{\rm nms} \quad \forall\, j<i,$$

and only masks with predicted confidence above a tuned $32\times32$0 are retained. No new network weights or loss functions are introduced. A common misconception would be to treat ParticleSAM as a retrained architecture; the paper is explicit that it remains a zero-shot or coarse-tuned adaptation of SAM’s ViT-B rather than a newly optimized segmentation model (Zhou et al., 5 Aug 2025).

3. Synthetic dataset generation and labeling pipeline

A central component of ParticleSAM is its automated data engine, which synthesizes dense multi-particle scenes from a library of isolated particle crops. This dataset is described as a benchmark for visual material quality control automation, and the pipeline is designed to provide both image synthesis and labeling without manual dense annotation (Zhou et al., 5 Aug 2025).

Library creation begins with a “Particle Segmentor.” The input consists of single-layer conveyor images $32\times32$1. A first SAM pass produces coarse masks $32\times32$2. These masks are then refined by extracting negative prompts at bounding-box corners and positive prompts at high-curvature points of $32\times32$3, and re-encoding those points into SAM’s prompt encoder to obtain refined masks $32\times32$4. Morphological closing and opening are then applied:

$32\times32$5

where $32\times32$6 and $32\times32$7 denote dilation and erosion with structuring elements $32\times32$8 and $32\times32$9. The output is a library of $64\times64$0 pairs, each tagged with a “farthest-pair” diameter

$64\times64$1

which assigns the particle to one of 8 size classes spanning 4 mm to 63 mm (Zhou et al., 5 Aug 2025).

Synthetic multi-particle generation then samples a total particle count $64\times64$2 from a prescribed Particle Size Distribution. The examples given are a uniform setting with $64\times64$3 for each class $64\times64$4, and a Gaussian setting with $64\times64$5. For each sampled particle, one library instance $64\times64$6 is drawn from class $64\times64$7 and transformed by

$64\times64$8

where $64\times64$9, $4096\times4096$0 denotes horizontal flip, and $4096\times4096$1 is color-jitter. Scaling is deliberately omitted in order to preserve real-world size. Placement on a $4096\times4096$2 background $4096\times4096$3 is constrained so that the visible fraction of each mask after occlusion lies in $4096\times4096$4. In multi-layer L3 images, a stacking order is also assigned, with small particles at the bottom and large particles on top (Zhou et al., 5 Aug 2025).

The synthesized image is produced by compositing all particles over the background:

$4096\times4096$5

where $4096\times4096$6 are the transformed binary masks in the selected layer order. The pipeline outputs a high-resolution RGB image $4096\times4096$7, a packed PGM mask image, and PSD metadata. The stages L1 to L3 vary occlusion, class mixing, and number of layers, ranging from none to heavy occlusion (Zhou et al., 5 Aug 2025).

4. Adaptation procedure and hyperparameter search

ParticleSAM does not retrain SAM’s weights. Instead, the paper describes a progressive hyperparameter search on an “adaptation set,” defined as a random subset of L1 and L2. The search schedule proceeds in two phases: Phase 1 tunes on L1, which has no occlusion, descending by class size; Phase 2 tunes on L2, which is occluded, again from large to small classes (Zhou et al., 5 Aug 2025).

The principal tuned parameters are the NMS IoU threshold, the confidence cutoff, the grid resolution, and the maximum area threshold used to remove conveyor-background masks. The reported operating values are approximately

• $4096\times4096$8,
• $4096\times4096$9,
• $\times5$0,
• and a task-specific $\times5$1 (Zhou et al., 5 Aug 2025).

No gradient-based optimizer such as Adam, and no learning-rate schedule, is involved. The paper therefore places ParticleSAM in a prompt-free, near zero-shot adaptation regime at deployment time. This suggests a methodological emphasis on inference-time retuning of a foundation model rather than data-intensive end-to-end retraining.

5. Reported evaluation results

The evaluation uses mean IoU and mean Average Precision at IoU threshold $\times5$2. The paper defines

$\times5$3

and states that $\times5$4 counts a prediction as correct if and only if $\times5$5 (Zhou et al., 5 Aug 2025).

Data	Model	Reported metrics
L1	SAM	mIoU 55.8%; mAP$\times5$6 53.6%; mAP$\times5$7 41.2%
L1	ParticleSAM	mIoU 68.2%; mAP$\times5$8 61.1%; mAP$\times5$9 48.5%
L2-l	SAM	mIoU 61.1%; mAP$64\times64$0 73.2%; mAP$64\times64$1 65.5%
L2-l	ParticleSAM	mIoU 85.5%; mAP$64\times64$2 89.6%; mAP$64\times64$3 85.7%
L3-m	SAM	mIoU 18.9%; mAP$64\times64$4 55.2%; mAP$64\times64$5 44.7%
L3-m	ParticleSAM	mIoU 69.3%; mAP$64\times64$6 82.8%; mAP$64\times64$7 76.9%

The paper highlights several observations. On heavily occluded mixed-class L3_h, ParticleSAM improves mIoU by approximately 26% absolute. It also reports that the drop from $64\times64$8 to $64\times64$9 is much smaller for ParticleSAM, which is interpreted as evidence of stable masks even at high IoU thresholds. At the same time, no statistical significance testing was reported, although the performance gains are described as large and consistent across splits (Zhou et al., 5 Aug 2025).

6. Qualitative behavior, failure modes, and broader relevance

The qualitative comparison emphasizes object separation and boundary quality. Figure 1 is reported to show that vanilla SAM often merges adjoining small stones or misses them entirely, whereas ParticleSAM recovers individual masks with crisper boundaries. Under heavy occlusion, both models degrade, but Figure 2 is summarized as showing ParticleSAM still recovering approximately 35% of stones versus less than 9% for vanilla SAM (Zhou et al., 5 Aug 2025).

The remaining failure cases are also specific. Extreme pile-up of sub-5 mm fragments remains difficult. The paper notes that sub-cropping could help in such cases, but at the cost of $t_{\rm nms}$0 latency, and also suggests that specialized small-object modules would help. This places a limit on the claim of general robustness: the adaptation improves dense small-object segmentation substantially, but does not eliminate breakdown in the most extreme scale and occlusion regime (Zhou et al., 5 Aug 2025).

Although developed for recycled construction aggregates, the method and its data engine are explicitly presented as applicable to domains with many small, densely packed or occluded objects, including cells in microscopy and grains in materials science. They are also framed as relevant when large-scale synthesized training or benchmark data with controlled PSD-like distributions are needed, and when real-time monitoring constraints make full fine-tuning impractical. A plausible implication is that ParticleSAM’s main contribution is not only a domain-specific benchmark result but a template for repurposing foundation segmentation models to tiny-object regimes through higher-resolution prompts, mask-area filtering, and domain-specific synthetic data (Zhou et al., 5 Aug 2025).