Adaptive Multi-prompt Contrastive Network for Few-shot Out-of-distribution Detection

We sincerely thank Reviewer oFtr for the valuable comments and provide the following detailed responses to all weaknesses.

---

W1: Figure 1 is not referenced or discussed in the main text. The caption of Figure 7 contains a error: "right" and "left" are mistakenly reversed.

Thanks a lot for your valuable suggestion. Figure 1 should be referenced in Introduction. Figure 1(a) should be placed in the first paragraph of Introduction. Figure 1(b) should be placed in the second paragraph of Introduction (Line 81). Figure 1(c) should be placed in the third paragraph of Introduction (Line 96). We will revise the caption of Figure 7 in our revised version.

W2: The paper heavily relies on CLIP for feature extraction and prompt engineering. While this is a common choice, it raises the question of whether using alternative models for feature extraction could yield different results.

Thanks a lot for your valuable suggestion. 1) Since all the compared methods utilize CLIP to extract features, we follow them to use CLIP for fair comparison. 2) For many other feature encoders, they include more complex calculations, which might make the designed OOD detector time-consuming. 3) The CLIP network can effectively extract the visual and textual features under the few-shot setting. Therefore, we utilize CLIP in our paper.

W3: The paper does not discuss the computational complexity or efficiency of the proposed method in comparison to existing methods.

We appreciate the reviewer’s comments regarding the computational complexity and efficiency of the proposed method. Compared to other methods, the training time and memory consumption are similar:

We evaluate the time and memory consumption of our proposed method compared with other baselines in above table and the results show that our proposed method is relatively compute-efficient. The evaluation is conducted on a single GTX-4090 GPU with a batch size as 64.

We sincerely thank Reviewer 8TNR for the valuable comments and provide the following detailed responses to all weaknesses.

---

W1: A little typo in Figure 1 caption. “(c) Brief framework of our method.” should be “(d) Brief framework of our method.”

Thanks a lot for your valuable suggestion. We will revise Figure 1 caption in our revised version.

W2: T-SNE visualization in Figure 3 is based on which dataset?

We appreciate the reviewer’s comments regarding T-SNE visualization in Figure 3. The dataset of T-SNE visualization in Figure 3 is based on the ImageNet-1k dataset.

W3: The sentence “As shown in Figure 3, different classes have distinct diversity” should be removed to the first paragraph of Section 3.2.

Thank you for your constructive suggestion. We will revise it in our revised version.

W4: In the paragraph above Equation (7), “Based on (7)”, is this Equation (7) or Equation (6)? Is it a typo?

Thanks a lot for your valuable comment. Yes. It should be Equation (6). We will revise it in our revised version.

W5: It would be better if authors could provide some failure examples of few-shot OOD detection.

In Few shot OOD, its paramount to generate ID and OOD prompts. Therefore, the dataset providing the labels (ID dataset) is key. It should be diverse enough. This is the case of ImageNet1K which is used in the benchmarks.

However, if we take a more specific dataset as ID, such as PLACES, which is less diverse, then the prompt engineering is less effective:

Here, the Places dataset (less classes) is treated as ID set and the ImageNet-1k dataset as OOD set (reversed compared to the benchmarks). We can find that when we switch the Places dataset and the ImageNet-1k dataset, the performance improvement is reduced.

We sincerely thank Reviewer J7Ff for the valuable comments and provide the following detailed responses to all weaknesses.

---

W1: Authors should move the definition of c to the third paragraph of Section 3.2.

Thanks a lot for your valuable suggestion. We will revise it in our revised version.

W2: In Eq. (7), $O \cdot \mathcal{M}_c^{pse}(t)$ should be $O \cdot \mathcal{M}_c^{pse}(t-1)$?

We thank the reviewer for the insightful suggestions. Yes. It should be $O \cdot \mathcal{M}_c^{pse}(t-1)$. We will revise it in our revised version.

W3: In Table 4, what is the fixed threshold?

We appreciate the reviewer’s comments regarding the fixed threshold in Table 4. It is the initial threshold. In the fixed setting, we do not update the threshold during training. Since there are different diversities in different classes, we want to learn different thresholds for different classes. Different diversities correspond to different distributions. In Section 3.3, we learn the distribution of each class in "Learning distribution" to obtain the corresponding diversity information. In "Intra-class distribution normalization", to fully learn the intra-class distribution of ID samples for better classification, we independently normalize the distribution for each class by $L_I^1$. In "Inter-class distribution normalization", we balance the distributions of all the classes by $L_I^2$.

W4: In Figure 4 and Figure 7, the sub-figures are too crowded.

We thank the reviewer for the insightful suggestions. We will redesign them in our final version.

We sincerely thank Reviewer 1bUn for the valuable comments and provide the following detailed responses to all weaknesses.

---

W1: Prompt learning is widely explored in the community. The idea of using learnable prompts for both ID and OOD classes is straightforward. However, the design of contrastive losses between different sets of prompts adds a meaningful contribution.

We appreciate the reviewer’s comments regarding learnable prompts and contrastive losses. Firstly, we combine $P$ learnable ID prefixes and the label name to generate the learnable ID prompts (LIPs). Also, we generate $S$ labelfixed OOD prompts (LFOPs) by introducing OOD labels from other datasets that are disjoint with the ID label set. Since the introduced OOD labels are often limited, we explore $Z$ label-adaptive OOD prompts (LAOPs) for each ID prompt. Besides, we align the image features and ID prompt features by a prompt-guided contrastive loss for ID classification.

W2: The authors should conduct a sensitivity analysis to understand the impact of these parameters on model performance.

Thanks a lot for your valuable suggestion. Due to the page limitation, we have analyzed some significant parameter in Figure 5. We are using hyperparameters $\tau_0, \ldots, \tau_5$ in the tool. We are reporting here the sensitivity analysis towards these hyperparameters on of the benchmark SUN, in the eight-shot setting:

W3: In the methodology section, there is a lack of discussion regarding the rationale behind certain technical design choices. For example, it is unclear how class-wise thresholding addresses the issue of sample diversity—this should be further elaborated.

We also appreciate the reviewer’s comments regarding the class-wise threshold. In the fixed setting, we do not update the threshold during training. Since there are different diversities in different classes, we want to learn different thresholds for different classes. Different diversities correspond to different distributions. In Section 3.3, we learn the distribution of each class in "Learning distribution" to obtain the corresponding diversity information. In "Intra-class distribution normalization", to fully learn the intra-class distribution of ID samples for better classification, we independently normalize the distribution for each class by $L_I^1$. In "Inter-class distribution normalization", we balance the distributions of all the classes by $L_I^2$. By learning the class-wise threshold and data distributions (intra-class distribution and inter-class distribution), we can obtain an adaptive classification decision boundary for each class, which effectively reduces the negative impact of different diversities on the challenging few-shot OOD detection task.

W4: the equation following Equation 3 is not numbered. Based on the context in the later sections, it seems to represent $L_1$, not $L_C$.

We thank the reviewer for the insightful suggestions.We will add the equation number in the final version. In fact, it is $L_C$, the $L_1$ is defined in Eq. (9).