Summary and Contributions: This paper proposes an algorithm for few-shot object detection based on the insight that hard negatives (namely badly localized boxes in the support set) are important. The proposed approach is shown to outperform prior art.
Strengths: - The insight that badly localized boxes are much needed negatives is a great one, and very important. - The results are impressive and clearly beat current statee-of-the-art. - The approach is clearly novel. - Few-shot object detection is a problem of great practical significance, so these results are important.
Weaknesses: (1) If the insight is that hard negatives are important, then a very simple baseline presents itself: why not take a simple object detector trained on the base classes, and simply finetune the detector head and bbox regressor head in the usual way on the novel classes? This would automatically use the badly localized examples. I am surprised that the authors did not include this baseline. (2)The numbers reported in Table 4 all use different network backbones. YOLO-FR uses DarkNet-19, Meta-Det uses VGG16, Meta-RCNN uses ResNet-101 (but without FPN or DCN). Meta-Det for e.g. pretrains the backbone on ImageNet. RepMet pretrains it on Coco. It is unclear what this paper pretrains it on. These differences are absolutely crucial and can easily explain any of the observed differences in performance. Ideally, the comparison would be apples to apples using identical backbones and pretraining regimes. But I accept that this may be too difficult. But at the very least, the paper should be explicit about exactly what the backbones and settings of each compared technique is. I cannot imagine accepting the paper without this, since without controlling for backbones and pretraining regimes, the conclusions are basically meaningless. (3) While this is not necessary for acceptance, I highly recommend the authors show results on Coco and/or LVIS, since they are much harder than Imagenet-LOC or Pascal VOC.
Correctness: See above.
Relation to Prior Work: Yes.
Additional Feedback: I want to see points (1) and (2) addressed for me to recommend acceptance. [EDIT]: Upgraded review based on rebuttal.
Summary and Contributions: The goal of this paper is to restore the information in negative proposals for few-shot object detection. The proposed method is built on the pipeline of RepMet by modifying and adding several modules to learn the embedding space and the representatives from positive and negative proposals using triplet losses. The main focus of the paper is on describing how the modifications can improve RepMet. To reduce the number of negative proposals, the paper proposes to use spectral clustering for choosing negative and positive representative proposals from the support images. The results show that the proposed method significantly outperforms RepMet on ImageNet-LOC and PASCAL VOC benchmarks for few-shot object detection.
Strengths: The writing of the paper is clear. The idea of including another branch in RepMet to learn negative embeddings from negative proposals seems to work well for object detection, since in object detection the issue of foreground-background imbalance is more essential and critical than in classification.
Weaknesses: The main idea introduced by the proposed method is to add negative representatives into RepMet. Although this modification seems to be effective, the novelty and contribution might be limited, because the problem that is addresed by this paper is specific for RepMet, due to the fact that RepMet does not take the negative proposals into consideration for learning the embeddings. Other few-shot methods like , MetaDet, and Meta R-CNN do not rely on representatives and hence do not suffer from this problem. It is reasonable and easier to understand why modeling the postive representatives is useful for few-shot object detection. However, the idea of learning 'negative-representatives' is a little strange since the diversity of background is much larger, and it might be less intuitive that a set of meaningful 'negative-representatives' can be learned from the hard negatives extracted from a few sample images and then can be used to distinguish the negative proposals in a query image that exhibits a very different background. It would be helpful if some examples of Cluster-Min can be visualized to provide some insights on what negative-representatives are learned.
Correctness: Because the backbone is pre-trained (using ImageNet or COCO), it is necessary to describe in detail whether the pre-trained data include samples of the novel classes. The RPN takes features from the backbone, and hence if the pre-trained backbone has "seen" some samples of novel classes, the underlying features could be more sensitive to the target foreground objects even if the RPN is trained from scratch. Although the experimental settings in this paper follow previous works like RepMet, MetaDet, and , the issue of data contamination through the pre-trained backbone should be discussed when evaluating few-shot learning methods. Triplet/ranking/margin losses are common in the literature of object detection and metric learning. It is practical to use the triplet loss for learning NP-embedding.
Clarity: Yes, the paper is well written.
Relation to Prior Work: The following NeurIPS paper on one-shot object detection is related and can be considered for comparison: Hsieh, T.I., Lo, Y.C., Chen, H.T., Liu, T.L.: One-shot object detection with co-attention and co-excitation. In: NeurIPS (2019) Although the experimental settings might be different, it seems that their method achieves better results on VOC dataset than the proposed method for novel classes. For the design of DML module, the following CVPR paper can also be referred to: Wang, X., Hua, Y., Kodirov, E., Hu, G., Garnier, R., Robertson, N.M.: Ranked list loss for deep metric learning. In: CVPR (2019)
Additional Feedback: Typos: L111: denotes --> denote L258: v.s. --> vs. L291-L293: add space between "e.g." and the number
Summary and Contributions: The work extends a distance-based few-shot detector by adding a new type of hard negative samples, which are regions poorly intersecting the same ground truth region as the positive sample (in addition to hard negatives of other classes). When applied to the RepMet detector, this leads to substantial performance boost.
Strengths: The work proposes a new principle in hard negatives selection for the triplet loss based learning of object detector with distance-based region classifier (e.g., RepMet). The proposed regions with 0.2<IoU<0.3 are declared as special type of hard negatives and a special attention is given to them by learning negative prototypes (representatives) for these regions. They are used in inference as well as the training. The approach is shown to achieve SOTA performance on two benchmarks, boosting the original RepMet (on Imagenet-LOC) and improving upon other detectors (on PASCAL VOC)
Weaknesses: Two obvious questions, that come to mind, were not answered in the paper. 1. The proposed method is not restricted to few-shot regime, it can be used to train a standard detector (which is one of RepMet modes). How would it rate in this scenario? 2. Why weren't the standard RepMet applied to the second benchmark (Pascal VOC), along with its extended version (the NP-RepMet)? I would expect to see these omissions filled (or explained, in case I am missing something).
Correctness: The claims, method and methodology, described in the paper seem correct.
Clarity: The paper is written with clarity and good language.
Relation to Prior Work: The relation to prior work is properly discussed.
Additional Feedback: I encourage the authors to test the NP-RepMet in the standard training regime, as the scope of proposed technique may be wider than just the few-shot scenario. Also I would like to see RepMet's performance on Pascal VOC (which is obviously within technical capacity of the authors). My comments were addressed in the authors response, I therefore maintain my score.
Summary and Contributions: In this paper, the authors have introduced a new idea of the restoration of negative information in few-shot detection.They proposed a new negative- and positive-representative based metric learning framework with negative information incorporated in different modules for better feature steering in the embedding space.
Strengths: - The authors have clearly presented the motivation of their method in Figure 1. - They have conducted extensive experiments to demonstrate the effectiveness of their approach.
Weaknesses: - For experimental results in Table 4, the authors did not report the results of  for 3-shot. From the results, can we conclude that the proposed method does not have better performance than ? The authors fail to provide a detailed analysis on this. - The authors take IoU>0.7 as positives and 0.2<IoU<0.3 as negatives. How does this parameter determined? This is not included in the current version. - The authors fail to compare against state-of-the-art hard negative mining methods in the literature, although they have not been directly applied to few-shot object detection.
Relation to Prior Work: Yes