W1,Q1,Q2: Motivation does not match methodology design properly.

The motivation and methodology design are related and clear (supported by sZmi,rhP3).

• The goal of In-Context Fusion is to establish the correlations between reference and target (see Line 163-164), understanding the complex information within the context, which is crucial for the in-context segmentation task (supported by sZmi).
• The proposed M-Former is designed to address task ambiguity in prompts, with the challenges for M-Former regarding task ambiguity analyzed in Lines 181-185.
• This paper discusses the relationships between different tasks in detail (Line 112-117) and unifies them using instance segmentation, allowing different tasks to use the same model. Therefore, the loss is reasonable (supported by nRMP).

The paper clearly states the motivation of all designs, and their effectiveness is verified through experiments. This has also been recognized by other reviewers, aligning our method's design with our motivation.

W2,Q3: Insights and relevance.

The contributions and insights are as follows (please see General Response for details):

1. Ambiguities in Visual Prompting
2. Investigating the Capabilities of VFMs
3. Lightweight and Effective Decoder

We believe the module designs are relevant to our motivation. We design M-Former to address task ambiguity. Information from different tasks can interfere, causing incorrect predictions (Lines 181-185). Mask2Former and DETR are not designed for in-context segmentation, making it difficult to address these challenges. M-Former's dual-path structure and shared SA enable effective information interaction and prevent task interference, balancing efficient in-context segmentation decoding and resolving task ambiguity in prompts. In Table 5c, compared to Mask2Former, M-Former effectively resolves task ambiguity, resulting in improvements across various tasks.

W3,W6,Q4,Q9: Transformer along with Hungarian algorithm is not uncommon in modern Transformer based pipeline. Eq.3 Hungarian loss.

Hungarian Loss: We do not claim the loss as a contribution. Please refer to the General Response or Lines 65-71 of the paper for our contributions. The Hungarian algorithm has been widely used[A,B,C], and our use of it aligns with these works, making it reasonable and not a weakness. The loss with one-to-many matching used by the traditional method[D,E] will introduce NMS, which is not suitable for Transformer based method.

Transformer design: We analyze task ambiguity challenges (Lines181-185) and design M-Former to address them. The dual-path design and Shared SA ensure efficient decoding and prevent information confusion across tasks of different granularity. Unlike previous methods, SINE improves performance with fewer parameters, demonstrating our method's novelty. See our response to Reviewer rhP3 for more details.

W4(1),Q5,Q6: SINE is weaker than other works on FSS.

Our experiments show that SINE achieves significant improvements across multiple tasks and benchmarks (supported by sZmi and nRMP) and demonstrates solid practical utility (supported by rhP3).

The table below shows the average performance of Table 1, SINE achieves the best performance. Therefore, SINE's improvements are significant.

W4(2),Q7: VOS.

See our response to sZmi‘s W3.

W5(1),Q8: Impact of fusion part in Table 5c.

Mask Pooling part does not contain any parameters. Learning in-context information is demonstrated in In-Context Fusion (see Lines 170-171). Mask Pooling is a necessary operation for SINE and cannot be removed. Hence, the corresponding ablation was not performed in Table 5.

W5(2),Q8: Studies of specific loss, phototype length.

Phototype: The number of prototypes is determined by the prompts (reference images and masks) and is not a settable hyperparameter. The number equals the total count of different semantic categories in the in-context examples.

Loss: Hungarian loss has been widely validated in object detection and segmentation[A,B,C]. We use this loss to align with Mask2Former and facilitate comparison (see Table 5(c)). Loss with one-to-many matching[D,E] is not suitable. Other hyperparameters, e.g., query dimension, are also aligned with Mask2Former to eliminate their impact.

W7,Q10: More visualization comparison results.

Fig.3 in attached PDF shows further visual comparisons. For video tasks, SINE reduces tracking failures from intersections, viewpoint changes, and occlusions. SINE addresses task ambiguity, preventing errors in semantic segmentation where SegGPT fails, as shown in the second set of comparison results in Fig.3(a). In real-world image segmentation, SINE exhibits better class generalization than SegGPT, matching the LVIS-92i comparison in Table 1.

Limitation 1:The reviewer believes that the discussion is not proper.

We believe the discussion is proper. We discuss the connections and differences between SINE and related works (SegGPT and SAM), and the limitations of SINE(Line478-494). We believe these discussions are necessary. If the reviewer finds any inaccuracies, please specify them in detail so we can further address them.

Limitation 2: Some potential negative impacts such as job displacement brought by generalist model should be discussed.

We do not foresee any obvious undesirable ethical or social impacts now. We believe that generalist models are powerful tools for efficient production and will create more job opportunities rather than causing job displacement.

[A] End-to-end object detection with transformers. ECCV2020.

[B] Per-Pixel Classification is Not All You Need for Semantic Segmentation. NIPS2021.

[C] Masked-attention mask transformer for universal image segmentation. CVPR2022.

[D] Mask r-cnn. ICCV2017.

[E] SOLO: Segmenting Objects by Locations. ECCV2020.

Thank you for your comments. Due to the rebuttal's character limit, we couldn't fully address your concerns. Below, we provide a detailed explanation.

No rigid defintion or formulation is given, especially to the level of task. Neither does the substantial relevance to the task-level ambiguity.

The formulation of SegGPT can be represented as:

$f(x_r, y_r, x_t) \rightarrow y_t,~~~ y_t \in \{task_1, task_2, task_3\}$

When the given prompt $(x_r, y_r)$ cannot precisely indicate a specific task, the prompt is ambiguous, as shown in Fig. 1 of the paper. When SegGPT performs a task (e.g., $task_1$) with an ambiguous prompt, it might incorrectly output the results of $task_2$ or $task_3$. This is an important and unexplored problem.

The formulation of SINE is as follows::

$f(x_r, y_r, x_t) \rightarrow \{ y_t^{task_1}, y_t^{task_2}, y_t^{task_3} \}$

By providing results for all tasks, SINE avoids incorrect predictions caused by prompt ambiguity. Fig. 4(a) in the paper presents a visual comparison.

The text still considers the object level learning. The two losses in the M-Former are still implemented on the object level.

We address the ambiguity between ID, instance, and semantic segmentation. These tasks can be converted into instance segmentation. Using instance/object level segmentation allows for unified loss forms across different tasks. M-Former introduces a dual-path decoder and shared self-attention with a mask (Fig. 2, top right), enabling effective information interaction and preventing task interference. This balances in-context segmentation decoding and resolves prompt ambiguity.

Our design unifies training for different tasks and prevents task information contamination during decoding. During inference, the trained network can perform different tasks, which we believe is an advantage of our method.

Lightweight and Effective Decoder

Compared to SegGPT's 300M training parameters, our model has only 19M. Without the encoder, Mask2Former has 21M parameters. Thus, our method is more lightweight.

We do not claim the loss as a contribution does not necessarily mean the devised representation learning pipeline is techniqually novel. Would the authors acknowledge a paper's novelty and significance, if it claims to use U-Net with minor modifications for another task / application such as industrial segmentation for the first time?

We believe the value of a paper lies in whether it provides academic insights to the research community. For example, Marigold (CVPR 2024, Best Paper Award Candidate) first demonstrated stable diffusion's significant generalization in depth estimation. However, Marigold made no changes to U-Net's structure or training strategy. This does not change our view that Marigold is an outstanding and inspiring work.

In addition, our core claim is that we are the first to identify and explore task ambiguity in prompts, rather than making minor modifications for in-context segmentation. We believe this will inspire other work. The General Response summarizes our contributions and insights.

Table 1, few-shot segmentation, on three out of six experiments, SINE is weaker than either Painter, SegGPT or recent ICLR works.

1. SINE outperforms Painter on all datasets.
2. SegGPT is trained on COCO and PASCAL and uses a Context Ensemble strategy for few-shot settings. SINE is not trained on PASCAL and is not designed for few-shot settings. However, SINE achieves the best performance on 1-shot in COCO-20i and PASCAL-5i, and is comparable to SegGPT in few-shot settings.
3. On LVIS-92i, SegGPT is only weaker than Matcher because Matcher uses SAM, and its larger training set better aligns with the LVIS.

Eq.3 Hungarian loss. Using the same loss as other works for fair evaluation is not a proper excuse to not study the impact of some common loss types. This aspect is unaddressed.

In the rebuttal, we explained that Hungarian loss is essential for Detection/Segmentation Transformer methods. These methods model detection/segmentation task as a set prediction problem, introducing one-to-one Hungarian matching. Traditional methods use a one-to-many matching for loss, which requires NMS and is not applicable here. Could the reviewer specify which loss should be used for comparison?

Limited visual results.

The visual results in the rebuttal are limited to the one-page PDF requirement. We will provide more visual comparisons in the paper.

Limitation: Efficient production definately leads to the loss of some old, tradtional and off-the-shelf jobs. This does not contradict create more job opportunities. Why is it difficult for the authors to acknowledge this aspect?

We will add this discussion in the paper. However, almost all AI model developments may cause such issues in the short term, but we cannot stop the advancement of AI because of this. We believe that, in the long run, the benefits of developing general models for human society outweigh the drawbacks.

Thanks the authors for the rebuttal.

Unfortunately, the rebuttal keeps repeating the comments and views from other reviewers, instead of seriously addressing the specific concerns raised by me. For example:

• Q1：I am not convinced from the text from the submission mentioned by the authors. The text still considers the object level learning. No rigid defintion or formulation is given, especially to the level of task. Neither does the substantial relevance to the task-level ambiguity.

• Q2: The two losses in the M-Former are still implemented on the object level. How do them relate to the task-level and the ambiguity? The rebuttal does not directly answer this question. Instead, it just keeps highlighting the view from other reviewers and mentions some void and lackluster general contributions.

• Besides, the authors in the rebuttal claim Lightweight and Effective Decoder. But the evidence in the main text supporting this aspect is not enough. The parameter comparison is only an ablation study, not with other types of decoders.

• Q3: The authors keep repeating the application value and insight in the rebuttal. However, how is the task ambiguity modeled, formulated and defined, and whether some theorical support can be found, are neither clarified. Besides, the concern The overall module designs are common in visual representation learning, which is less relevant to the task-specific guidance or the task-level ambiguity is unaddressed.

• Q4: We do not claim the loss as a contribution does not necessarily mean the devised representation learning pipeline is techniqually novel. The authors do not address ths aspect directly in the rebuttal. For example, would the authors acknowledge a paper's novelty and significance, if it claims to use U-Net with minor modifications for another task / application such as industrial segmentation for the first time?

• Q5 & Q7: I acknowledged that in some tables the proposed SINE shows state-of-the-art performance. However, my question also raises. Table 1, few-shot segmentation, on three out of six experiments, SINE is weaker than either Painter, SegGPT or recent ICLR works. The authors do not address this in the point-by-point response.

• I acknowledge from the rebuttal that the average performance is the best, but this does not contradict to the fact that, it shows significantly inferior performance in many of my listed settings. In fact, it further raises the concerns on whether it is stable or generalized enough.

• Q8 & Q9: Using the same loss as other works for fair evaluation is not a proper excuse to not study the impact of some common loss types. This aspect is unaddressed.

• Q10: Only limited visual results on limited datasets from SegGPT is provided. My concerns on more visual results from state-of-the-art methods are not addressed.

• Minor issue: Limitation: Efficient production definately leads to the loss of some old, tradtional and off-the-shelf jobs. This does not contradict create more job opportunities. Why is it difficult for the authors to acknowledge this aspect?

To conclude, my substantial concerns are not properly addressed. I keep my original rating and recommend clear reject this paper.

We believe that our response has answered the reviewer's concern. In addition, the reviewer's response at the last minute of ddl is irresponsible.

We thank you for your comments and the approval of our motivation and practical utility of SINE. We address your concerns here.

W1: The novelty of the overall idea and network structure is limited.

To address the reviewer's concerns, we first discuss the contributions and academic insights of our paper. Then, we discuss the differences between SINE and other methods (SAM, DETR, and Mask2Former). Finally, we outline some perspectives. We hope these responses adequately address the issues and concerns raised.

Contribution and Insight

In-context segmentation is an important proxy task for unifying different segmentation tasks, and it has garnered significant attention from the research community. Our paper delves deeply into this task, and we summarize our contributions, novelty, and research insights as follows:

1. Ambiguities in Visual Prompting: This paper is the first to explore the task ambiguity problem in prompts within in-context segmentation (supported by sZmi and nRMP). This is an important and novel issue that has been underexplored. We investigate the conflicts among multiple tasks in in-context segmentation from the perspective of task ambiguity and provide effective solutions, offering valuable insights to the research community.

2. Investigating the Capabilities of VFMs: Utilizing Visual Foundation Models (VFMs) to address various tasks is becoming a research trend. Our paper explores how to efficiently transfer the visual representations of VFMs to in-context segmentation, a research paradigm that currently lacks extensive exploration in the in-context segmentation field.

3. Lightweight and Effective Decoder: We thoroughly analyze the challenges brought by task ambiguity (see Lines 181-185) and design the M-Former structure to address these issues. The novel dual-path design and Shared SA in M-Former ensure efficient decoding while avoiding information confusion across tasks of different granularity. Additionally, unlike previous methods, SINE achieves significant performance improvements with fewer trainable network parameters (19M). Thus, our method also demonstrates novelty.

Differences between SINE and other methods

Based on the above contributions, SINE fundamentally differs from SAM, DETR, and Mask2Former:

Comparison with SAM

• SAM is a promptable segmentation model that takes a point as input and outputs the corresponding mask. The ambiguity lies in the uncertainty of the segmentation granularity represented by the point. SAM addresses this by introducing different queries in the decoder to obtain masks of different granularities.

• Resolving ambiguity in prompts (references) for in-context segmentation poses a more challenging task. In-context segmentation requires understanding the information in the prompt (e.g., category, position, shape) and learning the complex interaction relationships between the reference and target (approved by sZmi). Misunderstanding the prompt leads to incorrect outputs. Therefore, unlike SAM, SINE must learn the complex contextual relationships and avoid information confusion across different tasks.

Comparison with DETR and Mask2Former

• DETR and Mask2Former aim to train all model parameters on a specific dataset (e.g., COCO) to perform detection or segmentation on a limited number of categories.

• SINE aims to effectively perform in-context segmentation in the open world by using the off-the-shelf VFMs. The parameters of the encoder are frozen. We have analyzed the difficulties in addressing task ambiguity in in-context segmentation, challenges that do not exist in the general detection and segmentation task. Mask2Former and DETR are not designed for in-context segmentation, making it difficult to address these challenges. The dual-path design and shared SA in M-Former are introduced specifically to address these challenges. As shown in Table 5(c), compared to the Mask2Former decoder, M-Former effectively resolves task ambiguity, resulting in improvements across various tasks.

Perspective

Finally, we want to convey our perspective to the reviewer: In the era of large models, we cannot only focus on network architecture design. Different models with the same architecture can have vastly different characteristics (e.g., DINOv2 shows excellent patch-level matching ability, while CLIP excels in image-text retrieval). We should also focus more on how to fully leverage the potential of pre-trained models with fewer parameters and computations. From this perspective, SINE provides valuable insights to the academic community.

Thank you to the author for the rebuttal. My concerns have been partially addressed.

• "SINE must learn the complex contextual relationships and avoid information confusion across different tasks"

Compared with SAM, the contextual information seems heavily dependent on ground truth reference masks, whereas SAM can progressively segment objects in multiple turns. Other segmentation models, like those from co-segmentation, only require image groups as inputs. These examples make the claim confusing to me.

• "This paper is the first to explore the task ambiguity problem in prompts within in-context segmentation"

I agree with the author that we cannot solely focus on network architecture design in the era of large models. However, the insights presented appear to be only incrementally novel, as they mainly build on multiple previous works.

Thank you very much for your efforts and time in further discussing our work. Below, we provide a more detailed clarification regarding the remaining concerns.

Compared with SAM, the contextual information seems heavily dependent on ground truth reference masks, whereas SAM can progressively segment objects in multiple turns. Other segmentation models, like those from co-segmentation, only require image groups as inputs. These examples make the claim confusing to me.

We will address the reviewer's question from the following two points. If our explanation is unclear, we would welcome further discussion.

1. SINE is more cost-effective and broadly applicable. When using SAM for processing a large number of images, it requires human interaction to segment each image, leading to significant labor costs. Co-segmentation requires ensuring that a group of images contains objects with the same semantic conception and cannot segment multiple objects with different semantics in complex scenes simultaneously. This limitation hinders its widespread application. In contrast, SINE only needs a single in-context example containing a reference image and mask to handle target images in different tasks without human intervention, making it more cost-effective and broadly applicable.

2. Compared with SAM, why does SINE need contextual information. For each image, SAM segments objects based on human interaction without learning semantics. This allows SAM to progressively segment objects in multiple turns, but its drawback is the need for human input for every image, leading to significant labor costs. In contrast, SINE can batch-process images using just one image and its mask, a practical advantage that SAM lacks. To achieve this, SINE needs to understand the contextual relationship between the reference image and target images. In fact, SAM and SINE represent two vertically developed directions of segmentation foundation models and can complement each other. For instance, in auto-labeling, SAM could label objects (e.g., a dog) in the first image, and SINE could use that image as an in-context example to label subsequent images, reducing costs.

I agree with the author that we cannot solely focus on network architecture design in the era of large models. However, the insights presented appear to be only incrementally novel, as they mainly build on multiple previous works.

We appreciate the reviewer's agreement with our perspective. Below, we explain why our insights are not incremental:

In widely used segmentation tasks such as semantic and instance segmentation, traditional methods require assigning a category label to the object, which becomes challenging when scaled to thousands of categories. In-context segmentation overcomes these limitations by using visual concepts as annotations. However, ambiguous prompts in in-context segmentation can lead to incorrect predictions, limiting its practical use. We are the first to consider and resolve task ambiguity in prompts for in-context segmentation, a significant contribution in both research and application, and we believe this challenge is inherently novel.

In exploring the potential of DINOv2 for in-context segmentation, the challenge lies in how to maximize the reuse of the foundational model's capabilities across different tasks while still distinguishing the model's decisions for each task. This is a technical issue that cannot be solved by simply using methods like DETR. For example, as shown in Fig. 1 of the paper, an ID query for Prof. Geoffrey Hinton might get confused with broader information about the "person" category when using the traditional DETR decoder to process ID queries, instance queries, and semantic prototypes simultaneously. These technical challenges motivated us to propose the novel M-Former design:

• Dual-Path Structure: M-Former's input includes queries and prototypes. Queries handle instance-level predictions, while prototypes, derived from the reference image, maintain semantic-level category information. SINE processes them separately, forming a dual-path structure to enable query learning and avoid interference from prototype's semantic-level information.

• Shared SA: Queries include both ID and instance queries, representing different levels of granularity. To prevent interaction between ID queries and coarse-grained instance queries or prototypes, we introduce a well-designed attention mask (Fig. 2, top right) in the shared SA to effectively address this issue.

As shown in the table below, compared with recent method that directly adopt the DETR structure without considering task ambiguity, SINE achieves significant performance improvements.

[A] Visual in-context prompting. CVPR 2024.

Based on the considerations mentioned above, our insights are not incremental but provide valuable contributions to the research community.

Thank you to the author for their detailed analysis and explanation. In my opinion, the insights and novelty presented in the paper are limited, leading me to consider it as a borderline-to-rejection paper. I appreciate the extensive effort put into evaluating the proposed method. As a result, recognizing that the method is demonstrated as a strong baseline, I have adjusted my rating to borderline.

We thank you for your comments and the approval of our motivation and performance. We address your concerns here.

W1: Discussion of more ambiguity in visual prompts.

Thanks for your helpful suggestions. SINE is the first work to highlight task ambiguity in the visual prompts of in-context segmentation, initially focusing on resolving ambiguities among ID, instance, and semantic segmentation tasks (as these are more important and commonly used). We believe addressing ambiguities at this level is meaningful. For more complex ambiguities, such as full objects and parts, spatial position, category, and color, these can be addressed by incorporating multimodal in-context examples (e.g., image and text). We will add more discussions on this in the paper.

W2: Comparison of SINE and SegGPT by only using ADE20K, COCO.

The table below compares the one-shot semantic segmentation results of training SINE using only ADE20K and COCO with SegGPT. SINE outperforms SegGPT on three benchmarks. Notably, SINE achieves 10% higher mIoU than SegGPT on LVIS-92i, indicating stronger class generalization capability in real-world image segmentation compared to SegGPT. Additionally, we are the first to explore the use of Objects365 in in-context segmentation. Table 5(a) of the paper shows that SINE's effective design leads to further improvements in generalization capability with the inclusion of Objects365.

W3,Q1: Impact of different backbones.

We select DINOv2-S, DINOv2-B, DINOv2-L, and CLIP-L to explore the impact of different backbones. The conclusions are as follows:

1. DINOv2 Outperforms CLIP: DINOv2 achieves better performance than CLIP because it has general matching capabilities at both image and patch levels, allowing it to better understand complex contextual information between images. In contrast, CLIP captures image-text similarity, making it difficult to capture relationships between images, leading to poorer performance.

2. Larger DINOv2 Models Perform Better: Larger DINOv2 models have stronger representation capabilities, making it easier to capture contextual relationships, thus improving performance. This also indicates that SINE is scalable with the enhanced capabilities of the encoder.

Q2: How does SINE perform few-shot learning?

Multiple reference image features and masks are concatenated in the spatial dimension. The resulting feature and mask can be treated as a single reference image, and the subsequent process remains the same.

We thank you for your comments and the approval of our task formulation. We address your concerns here.

W1: Concerns on Generalizability.

Fig. 1 in the attached PDF shows SINE's capability in handling complex interaction relationships.

• In Fig. 1(a), the reference consists of multiple images, each containing different objects (box, cup, keyboard, mouse). When using these as in-context examples, SINE can segment one or more semantically different objects on a desk.
• In Fig. 1(b), with a reference containing only one object, the in-context example cannot represent complex interactions, and thus no segmentation result is provided.
• In Fig. 1(c), replacing multiple single-object images with a single image containing multiple objects yields the same effective results.

These experiments indicate that SINE has the potential to handle complex interaction relationships beyond semantically similar objects.

W2: Challenges with Complex In-Context Information.

We believe SINE can handle complex in-context infromation because the contextual relationships between complex scenarios can be captured more efficiently in the representation space of DINOv2. SINE leverages this characteristic of DINOv2 to accurately understand and fully capture intricate in-context information in multiple pairs of images. Extensive experiments show that SINE performs better in the few-shot setting than the one-shot setting, indicating its ability to effectively comprehend and capture in-context information from multiple samples. Figure 1(a) in the attached PDF also shows SINE's capability to understand and handle the complex relationships in in-context examples.

W3: Performance Discrepancy in Video Segmentation Compared to SegGPT.

We think the reason behind this performance gap is that SegGPT trains all model parameters (300M), while SINE uses a simpler in-context fusion module and fewer learnable parameters (19M). The detailed analysis is as follows:

SegGPT uses a broader dataset, concatenates references and targets spatially, and employs an ViT architecture. It trains all model parameters (300M), with self-attention effectively capturing relationships between video frames.

SINE aims to learn a general and lightweight decoder to efficiently transfer representations from pre-trained DINOv2 (encoders) to in-context segmentation. The frozen DINOv2 is limited to capture in-context information. For efficiency, SINE deploys only a simple in-context fusion module (1.58M) to learn in-context relationships, limiting its ability to handle inter-frame relations in complex videos.

Considering image tasks or simple videos (e.g., DAVIS), we believe that the overall performance of SINE are satisfactory. In paiticular, compared to recent generalist segmentation models SEEM[A] and DINOv[B] (which train all parameters), SINE's efficient design shows greater potential in video tasks (see the table below).

Although SINE currently has limitations in learning complex inter-frame relationships in videos, we believe that by designing more suitable In-Context Interaction module, the current paradigm holds greater potential for solving in-context segmentation tasks. We will explore this further in future work. These discussions will be added to the paper.

[A] Segment everything everywhere all at once. NIPS 2023.

[B] Visual in-context prompting. CVPR 2024.

W4: Unable to Handle Many Tasks SegGPT Supports.

SINE and SegGPT have different motivations.

• SegGPT aims to verify visual in-context learning can unify different segmentation tasks by using a broader range of datasets, such as semantic segmentation (ADE20K), instance segmentation (COCO), and part segmentation (PACO). Hence SegGPT can perform hierarchical in-context segmentation.
• SINE aims to resolve the task ambiguity in prompts. As the first work to highlight this problem, we initially focus on resolving ambiguities among ID, instance, and semantic segmentation tasks (as these are more important and commonly used).

When PACO is included as training data, SINE can perform part segmentation like SegGPT, as shown in Fig.2 of the attached PDF. This shows that our method does not overfit instance/semantic segmentation. Additionally, the results in Table 1 of the paper (LVIS-92i) and Fig.3 of the attached PDF indicate that SINE has stronger class generalization capability compared to SegGPT in real-world image segmentation.

Q1: Why is the target image the same as the raw image in Figure 4a ?

In Fig. 4(a), the giraffe demo is selected from two video frames with a large temporal gap. The reference is from an earlier frame with only one giraffe, while the subsequent frame gradually introduces another giraffe. This demo tests whether SINE can identify objects with the same ID without temporal information.

Q2: typos: L307 "Results Table 7 compares ...."

Thanks for pointing this out. It should be Table 4. We will fix this error in the paper.