Response to Reviewer fJQP
Thank you again for your valuable comments. We are committed to incorporating the suggested improvements into our paper, referencing **Q1-Q4**. We are confident that these updates will significantly enhance the clarity and overall readability of our work.
**Q1. Explanation of $A^ c \in\mathbb{R}^{C \times N}$, $A^c_{IP}$$\in\mathbb{R}^{C \times 1}$ and $A^{c}_{fro}$$\in\mathbb{R}^{C \times 1}$**
* **Simplify and clarify the notations: $A^c_{IP}$ and $A^c_{fro}$**
We notice the potential confusion regarding the notations $A^c_{IP}$, $A^c_{fro}$ and $A^{c}$ for the readers. To address this concern, we have taken steps to enhance clarity.
First, we will introduce additional clarifications at **Line187**: "$A^c_{IP}$ and $A^c_{fro}$ are specifically generated when computing the *self-distillation loss*. It is important to note that when processing an image through IP-CLIP without mask proposals, the resulting $A^c_{IP}$ is a matrix with dimensions ${C \times 1}$."
Besides, we will simplify $A^c_{IP}$ and $A^c_{fro}$ with $A_{S}$ (indicating the student net's score) and $A_{T}$ (indicating the teacher net's score), and showcase them in Fig. 2.
**Q2. Explanation of the final projection unit and** *w.o.* $B$
* **Add explanations of the final projection unit in Fig. 2**
We will enhance the clarity of the final projection unit by adding the following sentence in the caption of Fig. 2: "The final projection unit is an MLP module, used for reshaping the channels of $F_{feat}$ and $F_{cls}$."
* **The notation $w.o.B$ in Fig.2 will be revised as $w.o.M$**
We will revise "*w.o.* $B$ " with "*w.o.* $M$ ", as it corresponds to the process of generating $A^{c}_{IP}$, where the image is input to IP-CLIP without incorporating any mask proposal ($M$). The adjustment to "w.o. $M$ " better conveys the intended meaning and minimizes the potential for confusion.
In addition, the following explanation will be added into the caption of Fig. 2："*w.o.*$M$ denotes IP-CLIP Encoder processes image without utilizing mask proposals ($M$)".
**Q3. Explanation of the *mask-aware loss***
* **Clarify the normalization equation in Eq. 5**
At **Line180**, we will update the explanation to："We identify a discrepancy between the maximum values of $A^{c}$ and $S_{IoU}$. The maximum value of $A^{c}$ tends to approach 1, whereas the maximum value of $S_{IoU}$ ranges from 0.75 to 0.99. This inconsistency can hinder the alignment between $A^{c}$ and $S_{IoU}$. Therefore, we introduced a normalization technique for $S_{IoU}$ using Eq. 5."
* **Clarify the $\mathrm{smoothL1}$ function and the *mask-aware loss* in Eq. 6**
Eq. 6 provides the mathematical representation of the $\mathrm{smoothL1}$ function. To make it easier to understand, this equation will be re-formulized as:
$L_{ma} (A^c_{select}, S_{IoU}^{norm}) = \mathrm{smoothL1} (A^c_{select}, S_{IoU}^{norm}),$
Given two elements $x, y$ in $\mathrm{smoothL1}$. $\mathrm{If} ~ |x - y| < 1$, $\mathrm{smoothL1}(x, y) = 0.5\cdot (x - y)^2$; $\mathrm{Otherwise}$, $\mathrm{smoothL1}(x, y) = |x - y| - 0.5$.
**Q4. Explanation of starting mask attention from layer $L$**
We will refine the sentence beginning at **Line152** as follows: "We consider that CLIP's classification significantly relies on context information. In the first $L$ Transformer layers, the propagation of $F^{i}$ is the same as in standard CLIP. Specifically, $F_{cls}$ utilizes cross-attention with all pixels within $F_{feat}$, effectively retaining the context information."
****
We hope our responses have addressed all your concerns. If you have any further questions or comments, please kindly let us know, and we are happy to respond.