Segment Anything 2 (SAM 2) Overview

• Segment Anything 2 (SAM 2) is a foundation model for promptable visual segmentation that extends SAM’s capabilities to video by maintaining temporal state and propagating masks through sequences, using a hierarchical image encoder and memory attention.
• SAM 2 was developed for applications in camouflaged detection, underwater imaging, medical imaging, surgery, and remote sensing, outperforming baseline models and offering significant improvements in fine detail and context-dependent adjustments.
• One notable example of SAM 2’s application is in the field of surgical video segmentation, achieving high accuracy and efficiency with only a short promptable interaction.

Segment Anything 2 (SAM 2) is a foundation model for promptable visual segmentation in images and videos. It extends the original Segment Anything Model (SAM) by treating an image as a one-frame video and by maintaining temporal state through a streaming memory architecture. Given points, boxes, masks, or other prompts on one or more frames, SAM 2 generates object masks and propagates them through a sequence as a temporally consistent masklet. Its principal components are a hierarchical Hiera image encoder, prompt encoder, memory attention, mask decoder, memory encoder, object-presence head, object pointers, and a first-in-first-out memory bank (Ravi et al., 2024).

1. Development and task formulation

SAM 2 was developed to extend SAM’s promptable image segmentation paradigm to temporal visual data. The original SAM can segment arbitrary objects from prompts such as positive and negative points, bounding boxes, and masks, and it also provides an automatic mask-generation mode. It does not, however, maintain object identity or temporal state when applied independently to video frames. Independent frame processing can produce flicker, inconsistent object interpretations, loss of identity during occlusion, and the need to prompt or segment every frame separately.

SAM 2 formulates the task as Promptable Visual Segmentation (PVS). A video and one or more prompts are supplied on any frame; the model returns a mask on the prompted frame and a temporally consistent masklet, namely a sequence of masks corresponding to the target object. Prompts can be added on later frames as corrections without restarting propagation. A mask prompt on the first frame corresponds to conventional semi-supervised video object segmentation, whereas point and box prompts support interactive annotation and sparse initialization.

The model is trained on images and videos and is intended to segment arbitrary objects, object parts, and subparts rather than only predefined semantic categories. Its associated Segment Anything Video (SA-V) dataset contains 50.9K videos, 642.6K masklets, 35.5M masks, 4.2M frames, and approximately 196 hours of video. The dataset includes indoor and outdoor scenes from 47 countries, small and moving objects, occlusions, disappearances, and reappearances (Ravi et al., 2024).

The SA-V data engine iteratively combines model assistance and human verification. In its initial phase, annotators independently created masklets using SAM and manual tools, producing 16K masklets from 1.4K videos at approximately 37.8 seconds per frame. A second phase combined SAM initialization with mask-only SAM 2 propagation, producing 63.5K masklets at approximately 7.4 seconds per frame. A third phase used full SAM 2 with interactive corrections, producing 197K masklets at approximately 4.5 seconds per frame. The resulting workflow used human editing on 19.04% of frames and an average of 2.68 clicks per clicked frame.

The original SAM was trained on SA-1B, containing 11 million images and approximately 1.1 billion masks. SAM 2 combines SA-1B with SA-V manual data and internal video data, with the released mixture consisting approximately of 15.2% SA-1B, 70.0% SA-V, and 14.8% internal data. Training alternates between complete image batches and video batches.

2. Architecture and temporal memory

For a frame $I_t$, SAM 2 first computes unconditioned visual features using a Hiera image encoder. Memory attention then conditions the current-frame representation on previously stored spatial memories and object pointers. The prompt encoder embeds points, boxes, and masks. The mask decoder produces candidate masks, mask-quality or IoU predictions, an object-presence or occlusion score, and an object pointer. The memory encoder combines the current prediction with frame features and writes a compact representation to the memory bank. Processing proceeds one frame at a time.

The model can be summarized schematically as:

$$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$

For images, the memory bank is empty, so image segmentation is SAM 2’s single-frame mode. For videos, each frame is conditioned on prior prompted and unprompted frames.

Hiera image encoder

SAM 2 uses an MAE-pretrained hierarchical vision transformer called Hiera, with variants Hiera-T, Hiera-S, Hiera-B+, and Hiera-L. The default high-quality model is Hiera-B+ at $1024$-pixel resolution, while Hiera-L improves accuracy at lower speed.

Stride-16 and stride-32 features are fused through a feature pyramid network to form the principal frame embedding. Stride-4 and stride-8 features bypass memory attention and reach the mask decoder through skip connections, supporting small objects, fine structures, and thin boundaries. The encoder is run once per frame and produces unconditioned tokens; temporal conditioning occurs afterward through memory attention.

The image encoder removes relative positional biases and uses absolute-window positional encoding. Memory attention uses two-dimensional rotary positional embeddings and sinusoidal absolute positional embeddings for spatial memories. Object pointers do not receive rotary positional embeddings because they have no fixed spatial position.

Memory attention

Memory attention contains transformer blocks, with four blocks used by default. Each block applies self-attention to current-frame tokens, cross-attention from current-frame tokens to spatial memories and object-pointer tokens, and an MLP. In abstract form:

$$X_t^{(l+1)} = \operatorname{MLP} \left( \operatorname{CrossAttn} \left( \operatorname{SelfAttn}(X_t^{(l)}),M_t \right) \right),$$

where $X_t^{(l)}$ is the current-frame representation at layer $l$ and $M_t$ is the available memory.

The memory bank has two FIFO components:

• Recent-frame memories: memories from recent unprompted frames.
• Prompted-frame memories: memories from frames where a user supplied prompts.

The default ablations use six recent memories as a balance between accuracy and computational cost. Recent memories receive temporal position information, whereas prompted-frame memories do not because prompted frames may be widely separated in time. The bank also stores object pointers, providing semantic object identity complementary to spatial memory features.

SAM 2 stores memory features directly rather than passing them through a recurrent GRU. Ablations found that GRU memory was unnecessary and slightly less efficient. Reducing memory channels from 256 to 64 reduced storage by four times with little regression.

Prompt encoder and mask decoder

The prompt encoder is inherited from SAM. Positive and negative clicks are represented using positional encodings and learned embeddings indicating prompt type. Box corners are represented as sparse positional prompt tokens. Dense masks are embedded by convolutions and added to the frame embedding.

The mask decoder largely follows SAM’s two-way transformer design. It jointly updates prompt tokens and frame tokens, then produces masks through output mask tokens. High-resolution skip connections inject stride-4 and stride-8 features into the upsampling path.

SAM 2 produces multiple candidate masks when a prompt is ambiguous. A click on a part, for example, may correspond to the part itself, the whole object, or another visually coherent region. If later prompts do not disambiguate the alternatives, the candidate with the highest predicted IoU for the current frame is propagated.

The decoder also predicts whether the object is visible. This object-presence or occlusion head distinguishes an incorrect mask from a frame in which the object is genuinely absent, temporarily occluded, or outside the view.

Object pointers and memory encoding

The selected mask token is also used as an object pointer: a 256-dimensional semantic vector describing the object in the current frame. It is split into four 64-dimensional tokens before cross-attention. Object pointers improve long-term and difficult tracking, including re-identification after occlusion.

After prediction, the memory encoder downsamples the predicted mask and combines it with the unconditioned frame embedding:

$$\text{MemoryFeature}_t = \operatorname{Fuse} \left( E(I_t), \operatorname{Downsample}(P_t) \right).$$

The resulting spatial feature is projected to 64 channels and stored. The memory encoder reuses Hiera features rather than running a second image encoder.

3. Interaction, propagation, and training

SAM 2 supports initialization with a positive click, multiple positive and negative clicks, a bounding box, or a mask. A user can initialize an object on the first frame, propagate it through the video, and issue corrective prompts on later frames. Corrections are processed together with existing memory rather than restarting tracking.

This unified design combines prompt interpretation, segmentation, temporal propagation, correction, and memory update in one model. Multiple objects can be segmented, but each object is processed separately. Objects share the frame-level image embedding but do not exchange object-level information, which may make crowded multi-object scenes computationally expensive and limits explicit inter-object reasoning.

Training procedure

SAM 2 is first pretrained on SA-1B for approximately 90,000 steps at $1024\times1024$ resolution. The reported settings include AdamW, batch size 256, bfloat16 precision, learning rate $4\times10^{-4}$, weight decay 0.1, gradient clipping with maximum $$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$0 norm 0.1, seven correction points, focal mask-loss weight 20, Dice mask-loss weight 1, and IoU-loss weight 1.

Full image-video training lasts approximately 300,000 steps. Training alternates between image batches and video batches. The interactive prompting curriculum samples sequences of eight frames, randomly selects up to two frames for prompting or correction, uses at most three masklets per sequence, reverses temporal order with probability 0.5, and samples corrective clicks from ground-truth masks and model errors. Initial prompt probabilities are:

$$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$1

$$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$2

$$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$3

The training objective combines focal mask loss, Dice loss, IoU loss, and object-presence loss:

$$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$4

If a frame has no ground-truth mask, mask outputs are not supervised, but object-presence prediction remains supervised.

Evaluation protocols

Video evaluation uses the standard $$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$5 metric, combining region similarity $$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$6 and contour accuracy $$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$7. Image evaluation uses mean IoU. Interactive offline evaluation repeatedly selects the frame with the lowest IoU, adds three corrective clicks, and repropagates. Online evaluation pauses forward propagation when the current IoU falls below 0.75, adds three corrective clicks, and continues without revising already processed frames.

On nine dense zero-shot video datasets, SAM 2 obtained an average offline $$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$8 of 80.7 and online $$I_t \rightarrow \text{Hiera image encoder} \rightarrow \text{memory attention} \rightarrow \text{prompt-conditioned mask decoder} \rightarrow \text{mask, presence score, object pointer} \rightarrow \text{memory encoder}.$$9 of 79.7 after eight interacted frames, outperforming SAM combined with XMem++ and Cutie. The primary SAM 2 paper summarizes this as better accuracy with more than three fewer interactions than prior approaches.

On 17 semi-supervised video datasets, average $1024$0 was 64.3 with one click, 73.2 with three clicks, 75.4 with five clicks, 72.9 with a bounding box, and 77.6 with a ground-truth mask. On conventional VOS benchmarks, Hiera-L achieved $1024$1 on DAVIS17, $1024$2 on LVOS, and $1024$3 on the YTVOS19 $1024$4 metric.

For images, the Hiera-B+ model achieved 58.9 one-click mIoU on the SA-23 benchmark when trained on SA-1B and 61.9 with the full image-video mixture. At 130.1 FPS, it was approximately six times faster than SAM ViT-H at 21.7 FPS under the reported image-inference protocol (Ravi et al., 2024).

4. Performance characteristics and domain transfer

SAM 2’s performance is strongly conditional on prompt quality, object localization, and whether the task is prompted segmentation or autonomous object discovery. This distinction is central to evaluations across camouflage, underwater imagery, remote sensing, medicine, surgery, and industrial scenes.

Prompted versus automatic segmentation

On camouflaged-object detection, SAM 2 substantially outperforms SAM when prompted. With Shikra-generated coordinates, SAM 2 improves weighted F-measure on CAMO from 0.521 to 0.620, on COD10K from 0.482 to 0.565, and on NC4K from 0.570 to 0.672. With LLaVA-generated coordinates, the corresponding values are 0.633, 0.640, and 0.700. On MoCA-Mask video camouflaged-object detection, three points on the first frame yield $1024$5, $1024$6, and MAE $1024$7 (Tang et al., 2024).

The automatic results reverse this ranking. On CAMO, automatic $1024$8 is 0.606 for SAM and 0.184 for SAM 2. On COD10K, it is 0.701 for SAM and 0.271 for SAM 2. On NC4K, it is 0.696 for SAM and 0.251 for SAM 2. SAM 2 also generates approximately six to ten times fewer masks than SAM in these datasets. These findings indicate that strong prompted segmentation does not imply strong autonomous object discovery.

Underwater instance segmentation shows the same localization dependence. On UIIS, SAM 2 Hiera-L with ground-truth boxes achieves 70.6 mAP at 15.17 FPS, compared with 65.8 mAP at 3.89 FPS for SAM ViT-Huge. On USIS10K, Hiera-L with a ground-truth box reaches 77.2 mAP, whereas one-point prompting reaches 47.4 mAP. Dense automatic prompting with more than 900 points per image reduces speed to approximately 1.32–1.53 FPS depending on the Hiera variant (Lian et al., 2024).

A class-agnostic instance-segmentation study similarly found that SAM 2 is substantially stronger with ground-truth boxes than in automatic mode. For camouflaged instances, SAM2-L improves from 10.6 AP to 68.8 AP on COD10K and from 8.8 AP to 73.5 AP on NC4K when changing from automatic to ground-truth-box prompting. In salient-instance segmentation, SAM 2 also performs strongly with boxes but is generally weaker than SAM in automatic mode (Pei et al., 2024).

Medical imaging

In single-frame 2D medical segmentation, SAM 2 performs comparably to SAM across six 2D and twelve 3D medical datasets spanning MRI, CT, PET/CT, X-ray, and ultrasound. Box prompts generally outperform point prompts, and separate prompts for disconnected components are generally better than a single prompt. A reported example reaches IoU 0.907 for ilium segmentation in Xray-Hip, whereas MRI-Spine gray-matter segmentation reaches IoU 0.277.

In direct multi-frame 3D propagation, performance is much weaker. Treating volume slices as video frames produces an average result of approximately 0.19 IoU under the best reported configuration. Center-slice initialization is generally better than largest-object-slice initialization, bidirectional propagation is better than front-to-end propagation, and first-channel mask selection is often better than selecting the highest-confidence mask. The principal failure source is accumulated propagation error across slices (Dong et al., 2024).

Biomedical adaptation substantially improves performance. BioSAM 2 fine-tunes the image encoder and mask decoder on biomedical data while retaining SAM 2’s prompt and memory framework. On endoscopic instrument images, BioSAM 2 Hiera-T obtains DSC 0.6251 and NSD 0.6427. On microscopy images, it obtains F1 0.5792 and NSD 0.7436, compared with zero-shot SAM 2 Hiera-T F1 0.3566 under five clicks (Yan et al., 2024).

SAM-OCTA2 adapts SAM 2 to OCTA layer sequences using LoRA modules inserted into the image encoder. Only approximately 1.68% of parameters are trainable. It uses point prompts and propagates selected retinal vessels and the foveal avascular zone through ordered scanning layers. On OCTA-500, SAM-OCTA2 achieves FAZ Dice 0.9284 on the 6M field of view, but layer-sequence performance is substantially lower than en-face performance, with baseline sequence Dice values of 0.5487 for 6M retinal vessels and 0.6828 for 6M FAZ segmentation (Chen et al., 2024).

Surgical video

SAM 2 performs zero-shot surgical-tool segmentation from sparse prompts. On Endo Vis’17, a reported configuration obtains Dice 0.937, IoU 0.890, and MAE 0.018, outperforming the listed U-Net, UNet++, and TransUNet baselines. New tools entering the scene require additional prompts because they have no prior object representation in the memory bank (Lou et al., 2024).

On the SegSTRONG-C dataset, SAM 2 with five positive and five negative point prompts outperforms fully supervised models under background change, bleeding, and smoke. With one prompt on the first frame of a 300-frame sequence, it obtains DSC 0.9325 under background change, 0.8628 under bleeding, and 0.8660 under smoke. Low brightness remains a major failure mode, with DSC 0.4373, below all three fully supervised baselines (Shen et al., 2024).

Surgical SAM 2 introduces Efficient Frame Pruning, which removes redundant temporal memories using cosine similarity. With surgical fine-tuning, reduced resolution, and a ViT-Small backbone, the method reaches approximately 86 FPS compared with approximately 29 FPS for the evaluated vanilla SAM 2 configuration. Its reported inference memory is approximately 1.02–1.09 GB, compared with approximately 3.11–3.15 GB for vanilla SAM 2 on EndoVis17 and EndoVis18 (Liu et al., 2024).

Remote sensing and scientific imaging

For aerial solar-panel segmentation, SAM 2 provides only modest average gains over SAM but is more robust under sub-optimal lighting and point prompting. Across three unseen aerial datasets, SAM 2 with user boxes achieves mean IoU 0.75 and F1 0.85, whereas point prompting achieves mean IoU 0.51 and F1 0.62. At low resolution, point prompting falls to IoU 0.36 and F1 0.42 (Rafaeli et al., 2024).

SAM4EM uses SAM 2 for zero-shot single-particle tracking in liquid-phase transmission electron microscopy. It initializes particles with points or boxes, propagates indexed masklets, extracts centroids and orientations, and computes trajectory statistics. On synthetic LPTEM videos, it achieves $1024$9 of at least approximately 70%, compared with at most approximately 2% for the cited U-Net baseline (Goel et al., 6 Jan 2025).

5. Interpretability, memory control, and efficiency research

Several studies analyze or modify SAM 2’s temporal memory rather than its core segmentation representation.

An analysis of intermediate representations identifies five observational positions: raw input, image embeddings, memory-conditioned features, prompt-conditioned features, object pointer, and memory features. The object pointer is a 256-dimensional vector that most clearly separates target-present from target-absent frames and contains recoverable information about approximate object position and extent. Memory features are less object-specific because they retain broader scene context. Prompt attention is the stage at which visually similar distractors are most clearly suppressed (Bromley et al., 25 Feb 2025).

Det-SAM2 places YOLOv8 before SAM 2 to generate automatic box prompts. It adds online object insertion, bounded propagation, preload memory reuse, CPU/GPU offloading, old-frame release, FP16 storage, and asynchronous post-processing. The system targets arbitrarily long streams with bounded retained state. Its engineering estimates reduce naive repeated propagation from approximately $$X_t^{(l+1)} = \operatorname{MLP} \left( \operatorname{CrossAttn} \left( \operatorname{SelfAttn}(X_t^{(l)}),M_t \right) \right),$$0 frame-equivalent processing to approximately $$X_t^{(l+1)} = \operatorname{MLP} \left( \operatorname{CrossAttn} \left( \operatorname{SelfAttn}(X_t^{(l)}),M_t \right) \right),$$1 under cumulative buffering of size $$X_t^{(l+1)} = \operatorname{MLP} \left( \operatorname{CrossAttn} \left( \operatorname{SelfAttn}(X_t^{(l)}),M_t \right) \right),$$2 and maximum propagation length $$X_t^{(l+1)} = \operatorname{MLP} \left( \operatorname{CrossAttn} \left( \operatorname{SelfAttn}(X_t^{(l)}),M_t \right) \right),$$3. The report does not provide a standard segmentation benchmark or a direct numerical accuracy comparison with unmodified SAM 2 (Wang et al., 2024).

SAM2RL treats memory control as a reinforcement-learning problem. The SAM 2 image encoder, memory attention, memory encoder, and decoder remain frozen. A per-video PPO policy chooses whether to discard a new memory or replace one of six non-initial memory slots in a seven-entry bank. In an intentionally overfitted evaluation on 64 SA-V videos, SAM2RL obtains tracking quality 76.86%, accuracy 75.53%, and robustness 91.17%, compared with 71.95%, 73.01%, and 88.84% for SAM 2 (Adamyan et al., 11 Jul 2025).

Lean-SAM2 addresses memory attention and repeated full-frame encoding through three mechanisms: Target-Anchored Memory Pruning, Temporal Condensation with Insurance Memory, and Target-Anchored Risk-Aware Routing. It reports that memory attention plus image encoding account for more than 92% of latency on SAM2.1-Small, 83% on Base+, and up to 96% on Large in its LVOSv2 measurements. On LVOSv2 validation, it reports speedups of $$X_t^{(l+1)} = \operatorname{MLP} \left( \operatorname{CrossAttn} \left( \operatorname{SelfAttn}(X_t^{(l)}),M_t \right) \right),$$4 and $$X_t^{(l+1)} = \operatorname{MLP} \left( \operatorname{CrossAttn} \left( \operatorname{SelfAttn}(X_t^{(l)}),M_t \right) \right),$$5 on SAM2.1-Large and SAM2.1-Base+, with corresponding $$X_t^{(l+1)} = \operatorname{MLP} \left( \operatorname{CrossAttn} \left( \operatorname{SelfAttn}(X_t^{(l)}),M_t \right) \right),$$6 improvements of 5.0% and 3.6% (Ouyang et al., 22 Jul 2026).

The broader efficiency problem is that SAM 2 processes every frame through a high-capacity image encoder and attends to stored spatial memories. Reducing memory size, selectively routing image windows, condensing historical representations, or learning memory replacement policies can lower computational cost, but each method introduces trade-offs involving occlusion recovery, distractor suppression, temporal diversity, and long-range identity preservation.

6. Limitations, misconceptions, and research directions

SAM 2 is not an autonomous object detector merely because it can generate masks automatically. Its strongest capability is conditional segmentation after the target has been localized by a point, box, mask, detector, MLLM, or human. Automatic mask generation may produce too few candidate masks or fail to discover camouflaged, shadowed, low-contrast, or context-dependent objects.

Nor is SAM 2 a universally superior replacement for SAM. It is generally better for prompted image segmentation, video propagation, and sparse interactive correction, but several studies find that SAM is better in automatic image-mode discovery, especially for camouflage, salient instances, shadows, and other context-dependent concepts. The distinction between object localization and mask refinement is therefore essential.

SAM 2’s video memory is also not equivalent to native 3D understanding. Treating medical slices as video frames can be useful, but anisotropic spacing, abrupt anatomical changes, long volumes, slice ordering, and accumulated prediction errors create failure modes not encountered in ordinary video. Direct zero-shot volumetric propagation is generally unreliable without medical adaptation, bidirectional initialization, periodic correction, or redesigned memory.

Important failure modes include fast motion, long occlusions, shot changes, crowded scenes, similar objects, thin structures, low illumination, blur, bleeding, tool–surface interactions, small or low-contrast biomedical targets, vessel splitting across OCTA layers, and dense cell clusters. Multiple-object processing is independent, and object-level communication is not built into the released model.

The principal future directions implied by the research include:

• Prompt-free object discovery: integrating detectors, MLLMs, learned proposal generators, or task-aware adapters.
• Memory control: replacing recency heuristics with confidence-aware, target-anchored, learned, or reinforcement-learned memory policies.
• Long-video robustness: selective memory writing, occlusion-aware retrieval, object-presence monitoring, and bounded but recoverable history.
• Domain adaptation: fine-tuning visual encoders and mask decoders for biomedical, surgical, underwater, remote-sensing, and scientific imagery.
• 3D and volumetric segmentation: incorporating slice position, anisotropic spacing, bidirectional propagation, and explicit cross-slice consistency.
• Fine-detail preservation: improving high-resolution representation of thin structures, appendages, vessel branches, and instrument boundaries.
• Prompt robustness: evaluating imperfect human clicks, inaccurate boxes, noisy masks, prompt frequency, and detector localization errors.
• Evaluation methodology: separating localization, object discovery, mask quality, temporal identity, uncertainty, and computational cost rather than reporting a single segmentation score.

SAM 2 is therefore best characterized as a unified, prompt-driven image-and-video segmentation system whose defining capability is memory-mediated temporal propagation. Its performance is strongest when a target object is identified and remains sufficiently recoverable through time. Its principal limitations arise when the task requires autonomous object discovery, context-dependent interpretation, long-range memory management, fine-grained structural reasoning, or domain-specific semantics.