High-dimension Prototype is a Better Incremental Object Detection Learner

1. The quality and clarity of the figures need further improvement. The legend for Figure 1 is inconsistent, and Figure 3 is blurry and has a low resolution.
2. Some descriptions need to be more precise and clear. For example, in line 303, "the old class C_o" should refer to a "class set". The variable z in Equation (6) has not been introduced; it should refer to proposals generated from the ROI layer. Additionally, how are the means and variances of the K components of GMD calculated? Are they based on the features x? For incremental learning tasks, is the computation of these means and variances confirmed before or after training?
3. As mentioned in line 515, "this poses significant challenges to the network’s fitting capability.", it lacks an experiment on the selection of the number of Gaussian componenets for larger models.

1. Lack of discussion regarding method complexity and training time.
2. As the core contribution involves adopting Gaussian Mixture Distribution, it would be beneficial to conduct ablation studies on the number of components in both two-stage and multiple-stage incremental settings.
3. Minor: some of the best results are not boldfaced in Table 4.

1. Gaussian Mixture Distribution-based prototypes have been previously explored in class incremental learning[1-4]. Specifically, [1] highlights that embedding distributions in feature space are not convex or isotropic in practice. Thus, the author’s emphasis on the complexity of detection features compared to classification features may not provide substantial new insights. The innovation level is thus moderate.

2. This method is not specifically tailored for incremental object detection (IOD) tasks, which appears more suited for general class-incremental learning for modeling class distributions.

3. The integration of GMDP and DAPO at the base stage of the model may inherently enhance the detector’s feature representation capacity. Given that the upper bound of Faster R-CNN has been exceeded in certain settings (e.g., the 19-1 setting), the comparisons with existing methods may be skewed. They are not designed for incremental scenarios.

[1] Steering Prototypes With Prompt-Tuning for Rehearsal-Free Continual Learning.

[2] Contrastive Learning of Multivariate Gaussian Distributions of Incremental Classes for Continual Learning.

[3] Class-Incremental Mixture of Gaussians for Deep Continual Learning.

[4] Saving 100x Storage: Prototype Replay for Reconstructing Training Sample Distribution in Class-Incremental Semantic Segmentation.

1. The authors claim that their method learns higher-dimension prototypes. However, I cannot understand the definition of "high-dimension" here, does it mean that the representation of the ROI proposal feature vector $x$, where the prototypes are built on, is of high dimensionality? But as stated in the 306th-310th rows, the introduced dimension scaling progressive learning strategy transforms the original feature vector $x$ into a lower dimensional vector $z$, and calculates the prototypes based on $z$. This operation seems to conflict with the concept of "high-dimension", could the authors make more discussions about this?

2. In Eq. (3), the details of how to assign the ground-truth label $y_{i,c}$ for each proposal feature vector should be clarified, can we find the details of such label assignment operation in previous works?

3. From Table 5, it can be seen that the performance of the proposed method is quite sensitive to the number of Gaussian mixture components, which may increase the difficulty of tuning this hyper-parameter for different application scenarios.

4. In Section 3.2, there are some notations are confused. For example:

   1. In the 268th row, both the dimensionality of proposal feature $x_i$ and the size of input feature set are denoted as $N$.
   2. In the 276th to 277th rows, the component number of Gaussian mixture distribution is denoted as $K$, but it is also indicated by $M$ in Eq. (9) and (10).
   3. The notations of the proposal feature vector $x_i$ are bold in Eq. (4) and Eq. (7), but not in Eq. (2), (3) and (5). They should be consistent.