SAMST: Multi-Domain Methodologies

• SAMST is a context-dependent term used for distinct systems in football tracking, remote sensing segmentation, and style transfer, requiring careful disambiguation.
• In football tracking, SAMST combines SAM for precise mask initialization with CSRT for continuous tracking and occlusion recovery, achieving high accuracy under light and moderate occlusion.
• In remote sensing, SAMST employs a transformer-based framework with SAM-based pseudo-label filtering, yielding improved segmentation metrics such as mIoU around 70.8% on the Potsdam dataset.

Within the 2025 literature considered here, SAMST is not a single standardized method name. It denotes at least two distinct research systems built around the Segment Anything Model: a football player tracking pipeline in "Team-Aware Football Player Tracking with SAM: An Appearance-Based Approach to Occlusion Recovery" (Ranasinghe et al., 9 Dec 2025), and a semi-supervised remote-sensing segmentation framework in "SAMST: A Transformer framework based on SAM pseudo label filtering for remote sensing semi-supervised semantic segmentation" (Yin et al., 16 Jul 2025). A closely related capitalization, SaMST, appears in "Pluggable Style Representation Learning for Multi-Style Transfer" (Liu et al., 26 Mar 2025) and refers to a style-aware multi-style transfer network rather than a SAM-based segmentation or tracking method. This polysemy makes disambiguation essential in technical use.

1. Scope and nomenclature

The acronym is reused across domains, but the underlying formulations differ substantially in task definition, model topology, and evaluation protocol. In the football-tracking usage, SAMST means SAM + CSRT with Team-Aware appearance models. In the remote-sensing usage, SAMST names a Transformer framework based on SAM pseudo label filtering for semi-supervised semantic segmentation. In the style-transfer literature, the related term SaMST abbreviates Style-aware Multi-Style Transfer and is orthographically distinct, though it is easily conflated with the other two usages (Ranasinghe et al., 9 Dec 2025, Yin et al., 16 Jul 2025, Liu et al., 26 Mar 2025).

Usage	Domain	Core formulation
SAMST	Football video analysis	SAM + CSRT with team-aware appearance models
SAMST	Remote sensing	SAM pseudo-label filtering for semi-supervised segmentation
SaMST	Image style transfer	Style-aware multi-style transfer with pluggable style representations

This suggests that "SAMST" is best treated as a context-dependent acronym rather than a canonical architecture. A plausible implication is that citations should always be accompanied by task and paper title to avoid ambiguity.

2. SAMST as a football player tracking system

In football tracking, SAMST is a lightweight pipeline organized into four stages—Initialization, Continuous Tracking, Occlusion Detection, and Occlusion Recovery—with a simple high-level controller. Initialization is performed on the first frame $I_1$ from user clicks $\{p_i\}$ and team IDs $\{\tau_i\}$. For each player, SAM generates candidate masks, the highest-scoring mask is retained, a bounding box is derived, an appearance representation is extracted, and a per-player CSRT tracker is initialized. The framewise loop then updates CSRT, refreshes the appearance history, tests for occlusion, and invokes recovery when the occlusion duration exceeds $L$ (Ranasinghe et al., 9 Dec 2025).

The mask-to-box and tracker-update primitives are defined explicitly as

$$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$

and

$$(b_i^t, c_i^t) = \mathrm{CSRT}_i.\mathrm{update}(I_t).$$

Occlusion detection is threshold-based rather than learned. A track is marked lost when CSRT confidence drops, box area changes abruptly, or the box center moves implausibly between consecutive frames:

$$\mathrm{lost}_i^t = [c_i^t < \tau_c] \;\vee\; \left[\left|\frac{\mathrm{Area}(b_i^t)}{\mathrm{Area}(b_i^{t-1})} - 1\right| > \tau_a\right] \;\vee\; \left[\lVert \mathrm{Center}(b_i^t)-\mathrm{Center}(b_i^{t-1})\rVert > \tau_p\right].$$

The reported thresholds are $\tau_c=0.3$, $\tau_a=0.5$ for a 50% area change, and $\tau_p=0.2\cdot \max(w,h)$.

The architecture is deliberately hybrid. SAM is used primarily for precise initialization and recovery proposal generation, while CSRT handles continuous tracking. The paper characterizes this combination as practical under resource constraints, especially when visibility is continuous. That characterization is consistent with the measured stability under light and moderately crowded play.

3. Appearance modeling, recovery logic, and reported tracking performance

The football version of SAMST uses a team-aware HSV histogram model for re-identification. Appearance extraction is restricted to the upper 60% of the player mask, reflecting the assumption that jersey color is the most informative domain cue. The jersey mask is

$\{p_i\}$0

Normalized $\{p_i\}$1 and $\{p_i\}$2 histograms with 32 bins each are concatenated, stored in a sliding-window appearance template of $\{p_i\}$3 frames, and averaged to form $\{p_i\}$4. Matching uses the Bhattacharyya similarity

$\{p_i\}$5

with a declared match when $\{p_i\}$6 and $\{p_i\}$7 (Ranasinghe et al., 9 Dec 2025).

Occlusion recovery is stratified into three regimes. Light occlusion corresponds to overlap of at most two players with CSRT confidence still above $\{p_i\}$8, so no explicit recovery is needed. Heavy occlusion involves 3–5+ players, with partial CSRT failures but possible recovery within $\{p_i\}$9 frames. Long-term occlusion is defined as off-screen absence beyond 15 frames, where re-acquisition attempts almost always fail. Recovery first predicts a center $\{\tau_i\}$0, samples point prompts around $\{\tau_i\}$1, runs SAM on those prompts, and re-initializes CSRT if any candidate mask yields similarity above $\{\tau_i\}$2.

The reported evaluation covers processing speed, tracking accuracy, and robustness. On 36 clips from the DFL Bundesliga Data Shootout, each 30 s @1080p (1920×1080), 25 FPS, with a typical $\{\tau_i\}$3–15 tracked players, the method runs at 7.6–7.7 FPS with stable memory usage around 1880 MB on an NVIDIA GTX 1660 Ti 6 GB, Intel i7-10870H, and 16 GB RAM, using Python 3.10, PyTorch 2.8.0, OpenCV 4.12, and SAM vit_h (2.38 GB). Aggregate results report 100% tracking success in light occlusion, 90% in heavy occlusion, and 8.66% in long-term occlusion. BBox Stability remains near 0.989–0.991 across scenarios; Identity Switches are 0 in light occlusion and 1 in heavy occlusion. For heavy occlusion, 19 events yielded 2 recovery attempts and 1 success, giving ORR = 50%. For the single long-term off-screen event, off-screen re-acquisition = 0%.

These results define the paper’s central trade-off. The SAM + CSRT combination is reported to provide consistent performance across crowd densities, but it struggles with long-term occlusions where players leave the frame. The paper therefore recommends stronger re-identification mechanisms, and suggests multi-camera, motion-prediction, dedicated ReID networks, or graph-based association as extensions for more severe recovery settings.

4. SAMST as a remote-sensing semi-supervised segmentation framework

In remote sensing, SAMST is an iterative semi-supervised semantic segmentation framework that alternates between Supervised Model Self-Training and a SAM-based Pseudo-label Refiner. The supervised model uses a Swin Transformer backbone with UPerHead as the primary decoder head and FCNHead as the auxiliary head. Stage 1 trains on labeled data alone; subsequent iterations train on both true labels and refined pseudo-labels. Stage 2 applies the current model to unlabeled images and refines raw predictions with three modules: the Threshold Filter Module (TFM), Prompt Generation Module (PGM), and Label Refinement Module (LRM) (Yin et al., 16 Jul 2025).

The TFM stores the full softmax output $\{\tau_i\}$4 and applies class-specific thresholds $\{\tau_i\}$5. The filtered pseudo-label is

$\{\tau_i\}$6

Pixels below threshold are assigned 255 (ignore). This preprocessing step removes low-confidence pseudo-labels before SAM is queried.

The PGM then interprets each connected component $\{\tau_i\}$7 in the filtered mask as a separate object. For each component, it computes the minimal enclosing rectangle, expands it by $\{\tau_i\}$8 pixels to obtain a box prompt $\{\tau_i\}$9, samples positive points inside the component subject to confidence and neighborhood constraints, and samples negative points inside the box but outside the component with analogous confidence and neighborhood constraints. The LRM feeds $L$0 to SAM, hole-fills the resulting binary mask, and then decides whether each mask pixel should inherit class $L$1 or be set to 255 based on the original model prediction $L$2, the class probability $L$3, and thresholds $L$4 and $L$5. Refined labels from all regions are finally stitched into a full pseudo-label map.

Self-training uses a weighted composite loss,

$L$6

where $L$7 and $L$8 are cross-entropy terms over labeled and pseudo-labeled pixels respectively, with per-class weights $L$9 and $$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$0 to counteract distribution shift. The model is retrained for 80,000 steps per pseudo-label iteration. The reported implementation uses 8× NVIDIA A40 GPUs, batch size 4 images/GPU, AdamW, lr = $$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$1, weight decay = $$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$2, a linear decay schedule with 1,500-step warm-up, and a 1/32 labeled ratio of Potsdam (108 patches), with the remainder unlabeled. Experiments perform one pseudo-label iteration.

On the Potsdam dataset at the 1/32 label ratio, the paper reports mIoU 70.80%, OA 86.44%, and mF1 81.65% for SAMST, compared with 67.00% / 93.48% / 78.56% for the baseline, 60.82% / 80.57% / 74.15% for LSST, 64.89% / 83.83% / 75.37% for ST++, and 66.58% / 83.62% / 78.23% for ClassHyPer. The paper also notes large per-class F1 gains, including clutter +17.7 pp and building +1.2 pp. The ablation discussion is primarily qualitative: TFM removes low-confidence noise, PGM + SAM refines object boundaries and separates glued instances, and LRM reduces semantic mismatches and false positives.

5. Related but distinct capitalization: SaMST in multi-style transfer

A separate line of work uses the near-homographic acronym SaMST, standing for Style-aware Multi-Style Transfer. Although it is orthographically close to SAMST, it does not rely on Segment Anything Model. Instead, it decouples style modeling from style transferring through a Style Codebook (SCB) and a single style-aware generator. Each style is represented by a learnable vector $$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$3, with $$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$4, and test-time stylization is performed by feeding the relevant code into a fixed network. Adding a new style requires learning only one new vector $$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$5, while the network weights remain unchanged (Liu et al., 26 Mar 2025).

The generator consists of three Style-Aware Blocks (SABs) between a lightweight encoder and decoder. Each SAB applies three style-conditioned operations: Style-wise Convolution (SConv), Style-representation Adaptive Instance Norm (SRAdaIN), and Style-wise Channel Modulation (SCM). The stylization objective combines content loss, style loss, reconstruction loss, and geometric consistency loss, with reported weights $$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$6, $$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$7, $$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$8, and $$b_i^t = (x_{\min}, y_{\min}, w, h) = \mathrm{BBox}(m_i^t),$$9. Training uses MS-COCO content images resized to $$(b_i^t, c_i^t) = \mathrm{CSRT}_i.\mathrm{update}(I_t).$$0, 50,000 style images from WikiArt + Internet resized to $$(b_i^t, c_i^t) = \mathrm{CSRT}_i.\mathrm{update}(I_t).$$1, Adam, initial lr = 0.001, halving every 0.75 M iterations, for a total of 3 M iterations.

Quantitatively, the paper reports that SaMST achieves ArtFID 25.20, CF 0.538, GE+LP 1.515, 5.31 G FLOPs, 1.03 ms/img, 0.91 M total params, and 0.11 M OIP. It also states that adding new styles after freezing the backbone requires retraining only the corresponding embeddings and that each new style converges in approximately 3,000 steps in less than 1 min on a 3090 GPU. Because the acronym differs only by case, this framework is relevant mainly as a terminological caution: in bibliographic or search contexts, SaMST and SAMST can be confused even though they address unrelated problems.

6. Common patterns, limitations, and recurrent misconceptions

A common misconception is that SAMST identifies a single method family. The surveyed literature shows instead that the name is reused for distinct systems with different objectives: tracking under occlusion in football video, pseudo-label refinement in semi-supervised remote sensing, and, with the capitalization SaMST, compact style conditioning in multi-style transfer (Ranasinghe et al., 9 Dec 2025, Yin et al., 16 Jul 2025, Liu et al., 26 Mar 2025).

Another recurrent misconception is that a foundation model alone resolves the core task difficulty. The football paper explicitly shows that SAM-based initialization does not by itself solve long-term occlusion or off-screen re-entry: success is high under continuous visibility, but off-screen re-acquisition is 0% and long-term tracking success is 8.66%. The remote-sensing paper likewise does not use SAM as a standalone semantic segmenter; instead, it wraps SAM inside a thresholded, prompt-driven, label-consistency pipeline because pseudo-label quality depends on confidence filtering, connected-component decomposition, and semantic post-checking. A plausible implication is that, in both cases, SAM functions as a strong geometric prior—particularly for masks and boundaries—while domain-specific modules remain necessary for temporal association or semantic calibration.

The limitations reported by the two SAMST papers are also domain-specific. In football tracking, the main failure mode is extended absence and severe occlusion beyond the recovery window, motivating dedicated ReID or graph-based association. In remote sensing, several hyperparameters are dataset-specific, including $$(b_i^t, c_i^t) = \mathrm{CSRT}_i.\mathrm{update}(I_t).$$2, $$(b_i^t, c_i^t) = \mathrm{CSRT}_i.\mathrm{update}(I_t).$$3, $$(b_i^t, c_i^t) = \mathrm{CSRT}_i.\mathrm{update}(I_t).$$4, $$(b_i^t, c_i^t) = \mathrm{CSRT}_i.\mathrm{update}(I_t).$$5, $$(b_i^t, c_i^t) = \mathrm{CSRT}_i.\mathrm{update}(I_t).$$6, $$(b_i^t, c_i^t) = \mathrm{CSRT}_i.\mathrm{update}(I_t).$$7, and $$(b_i^t, c_i^t) = \mathrm{CSRT}_i.\mathrm{update}(I_t).$$8, and running SAM per connected region introduces additional computation. In SaMST, the reported limitation is different: the method cannot instantly adapt to a completely unseen style at test time because it requires training or incremental learning of the corresponding embedding.

Taken together, these works position SAMST not as a singular algorithmic object but as a recurring naming pattern for lightweight systems that combine a large prior—SAM in tracking and segmentation, or a compact learned style codebook in SaMST—with specialized downstream mechanisms. The unifying theme is not architectural identity but the use of a strong pretrained prior plus small task-specific components to improve robustness under limited computation, limited labels, or both.