Quality-Improved and Property-Preserved Polarimetric Imaging via Complementarily Fusing

Reviewer WQ4j

• Issues about the synthetic dataset.

  • The information can be found in Line215-223. Our synthetic dataset is generated from the PLIE dataset [32] (a dataset about polarized image low-light enhancement), which provides short-exposure ($T_{short}$) polarized snapshots that suffer from low-light noise (serve as $\mathcal{L}$) along with the corresponding long-exposure ($T_{long} \approx 10T_{short}$) high-quality reference snapshots (serve as $\mathcal{I}$) captured by a Lucid Vision Phoenix polarization camera with a tripod. We only generate the blurry polarized snapshots ($\mathcal{B}$) by ourselves.
  • As for the exposure time, since $T_{short}$ varies across the PLIE dataset [32] dataset, our synthetic dataset could cover a long range of different exposure time settings. As for the ratio between $T_{long}$ and $T_{short}$, since the PLIE dataset [32] fixes the ratio to about $10$, we cannot change it during the generating process of our synthetic dataset. However, we believe our method can still have the generalization ability to different ratios, since we have the blurry polarized snapshots as the input, which can directly provide the information about the ratio.
  • As for the motion blur, we have considered varying motion blur intensities. We adopt the approach proposed in [33] to generate the blurry polarized snapshots, which can generate different motion blur intensities and patterns. Besides, to generate more severe motion blur for increasing the diversity, we add impulsive variation [1] to the motion trajectories.
  • As for the sensor noise, since the images in PLIE dataset [32] are captured by a real polarization camera, they already include real sensor noise. Therefore, we do not need to add synthetic noise into the images.

• Issues about the reference images of real data.

  • See common issues 1.

• Issues about Figure 7.

  • We provide the comparison results from other methods in the attached PDF file (Figure B). We can see that the reflection-removed image with the fusing process of our framework contains more detailed textures and less reflection contamination than other methods.

  • In Figure 7, we feed a short-exposure noisy, a long-exposure blurry, and the fused polarized snapshots into a reflection removal network (NIPS19RSP [16]) respectively, and obtain the corresponding output images respectively. The label "reflection-contaminated input" denotes the scenes fed into the reflection removal method [16], and the label "reflection-removed output" denotes the output scenes from the reflection removal method [16] respectively.

  • The capturing process of the scenes in Figure 7 is similar to the real data shown in Figure 6, please refer to common issues 1.

• Issues about our claim.

  • Here we explain the differences between our claim and the papers mentioned in the reviews separately:
    • The claim of "the first time applying a fusing strategy to polarimetric imaging" in Line56 of our paper is about polarimetric imaging. Note that the term "polarimetric imaging" means outputting high-quality polarized snapshots captured by polarization cameras. Since existing methods designed for such a goal do not use the complementarily fusing strategy, our claim could be accurate.
    • The term "polarization image fusion" in the paper named as "Semantic-guided polarization image fusion method based on a dual-discriminator GAN" is the process of fusing an intensity image and a polarization parameter image solved by Stokes vector into a more detailed image. It means outputting unpolarized images with more details, which is about increasing the quality of intensity images.
    • The term "attention-bridged fusion" in the paper named as "Polarization-driven semantic segmentation via efficient attention-bridged fusion" is fusing polarization information into the semantic segmentation procedure, which is about network designing.

Reviewer SSTW

• Why not choose the LLCP dataset as the source data?
  • This is because the quality of PLIE dataset [32] used in our paper could be better than the LLCP dataset [25]. For example, overexposed regions often appear in the reference images of the LLCP dataset [25]. Training a network with such data will reduce its generalization ability.
• Can the proposed method be used to process the data captured using polarizers?
  • Yes, it is theoretically feasible, as long as we can obtain four polarized images of the same scene with different polarizer angles ($0^{\circ}, 45^{\circ}, 90^{\circ}, 135^{\circ}$). However, using polarizers can only capture a single polarized image in a single shot, which is less convenient compared with using polarization cameras.

Reviewer 4C9V

• Issues about the PSNR and SSIM values of the DoP and AoP.
  • PSNR is highly sensitive to small changes in pixel values, whereas SSIM considers structural information and spatial relationships within the image. Our method leverages the clean information from the blurry input, effectively reducing the noise level in the DoP and AoP, particularly in high-value regions, leading to higher PSNR scores compared to other methods. However, since the overall pixel values of the DoP and AoP are typically very small (much smaller than those of the images), structural details are usually significantly degraded (due to distortions in the low-value regions) and challenging to correct. In such situations, neither our method nor the compared methods can significantly improve the structural details, resulting in similar SSIM scores. Consequently, the improvement in SSIM is not as pronounced as in PSNR.
  • Generally, an improvement in PSNR can be positively correlated with the quality of downstream tasks. This is because higher PSNR indicates better preservation of polarization information, which can provide more accurate polarization cues for downstream tasks. However, the quantitative relationship between them cannot be precisely determined, as the quality of downstream tasks significantly depends on the specific methods designed for those tasks.
• Issues about the reference images of real data.
  • See common issues 1.
• Visual comparisons of the ablation study.
  • We take the scene in Figure 5 as an example, which can be found in the attached PDF file (Figure A). We can see that although there is not much difference compared to other ablation items (due to their similar quantitative scores), there is still a slight advantage in details such as cleaner background.

Thank you very much for pointing out this issue. We will update Figure 4 by replacing $\theta$ with $2\theta$ and revise the corresponding text once the camera-ready submission process reopens after the conference.