Weakly-Supervised Affordance Grounding Guided by Part-Level Semantic Priors

We are grateful to the reviewer for the valuable insights and suggestions. In the following, we provide point-by-point responses to each concern.

Marginal improvements of the modules (Weakness 1)

The performance gain brought by the combination of the three modules is indeed relatively modest. However, we would like to emphasize the complementary nature of the proposed modules. As discussed in Section 4.4, the reasoning module primarily improves the performance on the unseen split (due to enhanced generalization ability), the label refinement stage is particularly beneficial for the seen split (as the refined labels directly contribute to a better understanding of seen object categories), and the alignment loss benefits both splits (due to the robustness brought by introducing exocentric images). From the results in Table 2, each module demonstrates its intended effect, supporting the validity of our design. Moreover, by employing all three modules together, the model can achieve consistent and significant improvements over our baseline model across both splits. Therefore, we believe the design of these three modules is well-justified. As a whole they can make full use of the hints from the exocentric images, the textual input, and the pseudo labels.

Effects of foundation model aided alignment stage (Weakness 2 & Question 2)

We provide in-depth ablation study on the alignment process in Appendix D.3, and the results are shown in Table 8. We copy the table here for ease of reference.

By examining the effect of using the exocentric object masks ($M^{\text{exo}}_{\text{obj}}$) in the alignment process, we found that the cross-view feature alignment is not helpful without the masks. Focusing on the object area has a significant advantage over globally pooling the exocentric feature maps, since it excludes the irrelavent information from the alignment process.

In Appendix D.3, we have also conducted ablation studies on the design of the reasoning module, the post-processing strategy of the refinement stage, the choice of the encoder (based on Reviewer 7PPq's suggestion), and the choice of the VFM (based on Reviewer Yqjp's suggestion). Please let us know if there are further concerns.

Need of exocentric images (Weakness 3)

The proposed method indeed relies on exocentric images. However, we believe that unlabeled exocentric images, which do not need to be strictly paired with egocentric images (i.e., they may have different object instances), are relatively easy to obtain, especially when compared to pixel-level affordance annotations. For example, such images can be sourced from the internet (by searching for " $affordance category$

$object category$ "), video datasets capturing human activities, or generative models.

Relation with cross-view alignment in feature learning (Question 1)

To the best of our knowledge, alignment between multiple views is a common strategy in self-supervised learning (e.g., SimCLR[1], DINO[2]). These works leverage the consistency between different augmented views of the same image to learn high-quality visual features. In the context of our task, however, the egocentric/exocentric views (as defined by Cross-View-AG[3]) are unrelated to augmentations. Instead, they refer to two distinct images containing objects of the same category but in different states. Consequently, the view alignment in feature learning is fundamentally different from our alignment process. While the former focuses on consistency under perturbations, the latter aims to transfer specific knowledge from the exocentric view to the egocentric view.

[1] A simple framework for contrastive learning of visual representations. ICML 2020

[2] Emerging Properties in Self-Supervised Vision Transformers. ICCV 2021

[3] Learning affordance grounding from exocentric images. CVPR 2022

We hope the responses above have addressed the reviewer's concerns, and we are always open to further dicussions.

We sincerely appreciate the reviewer's constructive and thorough feedback. Below, we provide detailed responses to the reviewer's comments.

Failure cases analysis and limitations (Weakness 1 & Question 1)

We present some examples of the failure cases in Figure 12 in the Appendix. As detailed in the caption of Figure 12, we observe two kinds of typical failure cases. First, for some small or intricate affordance regions (e.g. the part of a suitcase that affords dragging), the pseudo labels, even after refinement, are not accurate enough. Second, for some objects in the unseen test set that exhibit significant differences in shape or structure compared to the training objects (e.g., holding a cup vs. holding a wine glass), the model's generalization ability still requires improvement.

We kindly refer the reviewer to Appendix E for our discussion on limitations and future directions. In brief, possible improvements include: incorporating external knowledge to enhance reasoning capabilities on novel categories, introducing finer-grained image correlations to improve the alignment module, scaling up the dataset size, and establishing a more principled taxonomy for affordance.

Problem of the label refinement stage (Weakness 2)

The reviewer is correct that the label refinement process may meet difficulty when faced with complex affordances and interactions. In the refinement stage, we aim to focus on the affordance categories that are related to human body, especially hands. Exocentric images of these categories (e.g. hold, hit, drag) usually contain clear human-object occulusions that are informative for our model. It is also worth noting that hand-related affordances hold particularly significant practical value for current research in embodied intelligence, as they are directly related to object manipulation.

Ad-hoc part name mapping (Weakness 3)

As stated in Line 240-241, the manually constructed part name mapping is sufficient for the experiments on AGD20K, while implementing an automatically generated mapping is also feasible. In Table 5 (in Appendix), we present some examples of such a mapping generated by GPT-4o and the prompt we use. The results align well with the handcrafted mapping.

Sensitivity to VFMs and effect of the refinement stage (Question 2)

We would like to address the sensitivity to VFMs from the following perspectives.

(1) It is important to choose the right type of VFMs in the first place.

In Appendix B.1, we explore three different ways of using VFMs for pseudo label generation, including directly using a multimodal LLM, directly using a foundation model specialized for detection, and combining a detection model with label mapping. From the results in Figure 6 and 7, we observe that only the third approach can produce reasonable labels in most cases. Therefore, the first step in using VFMs is to design a proper pipeline to translate its semantic priors into affordance-related information.

(2) The errors in the initial pseudo labels can be fixed.

Given the relatively high-quality pseudo labels, our approach can correct the errors within them through two mechanisms, as shown in Figure 9 and 10 in the Appendix. On one hand, the supervised training pipeline will fix some occasional errors. Though the pseudo labels may miss some instances or deviate from the correct region, the trained model can make correct predictions because a lot of samples in the same category have correct labels (Figure 9). On the other hand, the label refinement stage will deal with some systematic errors, i.e. the cases where the VFM produces incorrect labels for the majority of images within certain category. In Figure 10 we visualize the progress made by the refinement stage. These results effectively demonstrate the robustness of our approach to the errors of VFMs.

(3) The overall performance will increase with the capability of the VFMs.

Our pipeline is not tied with VLpart and SAM, and they can be replaced by similar models. To validate this, we perform an additional ablation study on the choice of VFMs for label generation.

Here, VLpart+SAM is our default choice. FastSAM [1] is a YOLO-based lightweight segmentation model which claims 50× higher speed than SAM. PartGLEE [2] is a very recent work following the task setting of VLpart. It can be observed that our approach achieves better performance with stronger foundation models. Specifically, using PartGLEE+SAM, the model establishes a new state of the art. Even when using the weaker VLPart+FastSAM, the model still outperforms previous CAM-based methods. We believe this is an encouraging sign, indicating that the development of general vision models will continue to drive progress in the field of affordance grounding.

This experiment will be added to Appendix D.3 in our revised version.

[1] Fast Segment Anything. https://arxiv.org/pdf/2306.12156

[2] PartGLEE: A Foundation Model for Recognizing and Parsing Any Objects. ECCV 2024

Computational cost analysis (Question 3)

We provide an analysis on the efficiency of our model in Appendix A.4, and the model statistics are listed in Table 3 and 4. In brief, our model has comparable inference speed (~49 fps) with previous methods like Cross-View-AG (~52 fps) and WSMA (~30 fps), and the computation is mainly concentrated at the visual encoder.

As for training, it takes about 1h to generate the initial pseudo labels for the training set. The label refinement stage requires approximately 1h, followed by around 3h for supervised training. The whole training process can be performed on a single NVIDIA GeForce RTX 2080Ti. As reference, LOCATE's training scheme takes about 7h in the same environment, while WSMA takes about 2h. So the training cost of our method is basically at the same level with previous methods. Besides, our supervised training process is more straightforward than previous works, involving neither the clustering operations used by LOCATE nor the non-negative matrix factorization employed in WSMA. (This paragraph will be added to Appendix A.4 in our revised version.)

We hope the responses above have addressed the reviewer's concerns, and we are always open to further dicussions.

We sincerely thank the reviewer for the time and effort, and we are very grateful for the constructive and positive review. In response to the reviewer's suggestion, we have performed an additional ablation study on the visual encoder.

For each encoder, we use the ViT-B version to make a fair comparison. It can be observed that all the evaluated encoders achieve performance surpassing previous CAM-based methods. Specifically, the default choice (CLIP) slightly outperforms DINO and OWL-ViT. This could be attributed to the use of CLIP encoder in the text branch, which enables the affordance query to interact more effectively with the visual features in CLIP's latent space. SAM, on the other hand, exhibits a larger performance gap compared to CLIP. One possible explanation is that SAM's encoder does not incorporate a class token, which hinders the cross-modal fusion (detailed in Section 3.2 and Appendix A.1) from extracting high-level semantic information. Conversely, DINOv2 achieves better results than the default option, likely owing to the favorable properties emerged from self-supervised representation learning. This also aligns with DINOv2's strong performance in dense recognition tasks.

In summary, our pipeline is well compatible with different visual encoders. As our goal is to establish a general framework that leverages foundation models for the affordance task, selecting the optimal visual encoder falls beyond the scope of this paper. However, we do observe that more advanced encoders yield superior overall performance, and we appreciate the reviewer's insightful remarks in this regard.

We sincerely appreciate the reviewer for providing a prompt and thorough emergency review. Here is our response to the reviewer's concerns.

Clarification of the baseline

The "baseline" method in Table 1 refers to supervised training with the initial pseudo labels generated by the foundation models, i.e., the method described in Section 3.3. We would like to make it clear that the strong baseline itself is part of our contributions. Its improvement over previous methods is primarily attributed to the shift from CAM-based prediction to supervised training using pseudo labels. As mentioned in Line 77-84, the CAM is focused on the most discriminative part of the image, while it may fail to accurately capture the affordance region associated with an action. In contrast, the pseudo labels generated by the foundation models are of higher quality, and enable a consistent pipeline for training and inference (as shown in Figure 1(c)).

Improvements of the proposed modules

In addition to the baseline method, the other part of our contribution lies in the design of the three extra modules. We would like to emphasize that these modules have necessary and complementary effects, as detailed in Section 4.4. The reasoning module is designed for improving generalization. Thus, as observed by the reviewer, it does not show advantage on the seen split (better SIM but worse KLD and NSS), where the model has encountered all kinds of objects during training and the test set does not demand cross-category generalization. On the unseen split, where generalization ability is essential, including the reasoning loss consistently enhances the performance on all metrics. Similarly, the refinement stage has more significant effect for the seen split, since the refined labels directly contribute to a better understanding of seen categories. The cross-view alignment works well for both splits, showing the benefits of introducing exocentric images to guide feature learning.

Also, from a more quantitative perspective, the improvements achieved by incorporating the three modules into the baseline model (e.g., -0.05 KLD on the Seen split and -0.10 KLD on the Unseen split) are comparable to the improvements brought by prior methods (e.g., Cross-View-AG+ vs. Cross-View-AG, WSMA vs. LOCATE). Overall, we believe that the design of these modules is necessary and meaningful.

We hope the content above addresses the reviewer's concerns regarding our experimental results, and we are always open to further dicussions.