PM-VIS: High-Performance Box-Supervised Video Instance Segmentation

In order to mitigate the noise in the generated pseudo masks, we employ HQ-SAM, which incorporates a learnable high-quality output token, to produce these pseudo masks, referred to as HQ-SAM-masks. The HQ-SAM model, as shown in Fig. 1 (HQ-SAM), significantly reduces the number of outrageous errors compared to SAM. Besides, as depicted in Fig. 2, by calculating the IoU between HQ-SAM-masks and gtmasks, we find that HQ-SAM-masks exhibits 6.46% higher high-quality (IoU greater than 0.9) instance count compared to SAM-Masks and has a lower instance count across all intervals of IoU less than 0.9 compared to SAM-Masks.

We observe that HQ-SAM-masks still exhibits diminished quality in scenarios involving occlusion and rapid motion. Taking into account the effectiveness of the IDOL algorithm in handling occlusion, we introduce IDOL-BoxInst-masks, a set of pseudo masks generated by the box-supervised VIS model IDOL-BoxInst. It combines box-supervised IIS model BoxInst [12] with VIS model IDOL [2] while excluding mask losses [16, 17] from IDOL. This model is trained on the dataset without mask annotations.

Despite aforementioned techniques, pseudo masks generated by weakly supervised model still struggle to guarantee the accuracy and completeness of object predictions, especially in scenes with subtle color variations. We observe that within HQ-SAM-masks and IDOL-BoxInst-masks, despite the presence of some subpar segmentations, there are many high-quality instance segmentations that reach the ground-truth level. Subsequently, we employ the semi-supervised Video Object Segmentation (VOS) model DeAOT [18] with initialization of object annotations that comes from the high-quality pseudo masks in IDOL-BoxInst-masks, to track the pseudo masks of the target throughout the entire video, resulting in a novel pseudo mask collection called Track-masks.

During the inference stage of the IDOL-BoxInst, the model predicts multiple masks for each real instances. However, it is challenging to ascertain the quality of each predicted mask. Using predicted target labels or IDs alone does not uniquely establish the relationship with instance gtboxes. To establish correspondences between predicted masks and gtboxes, we propose a method that utilizes both the IoU between gtboxes and predicted boxes, as well as the IoU between HQ-SAM-masks and predicted masks to determine their associations. Specifically, we utilize HQ-SAM-masks for temporal continuity and object matching, constraining the predicted pseudo masks using our proposed SCM (Scoring Corresponding Mask) confidence scores.

Upon closer examination of these segmented instances, two distinct issues within the segmented instance annotations appear: (i) overlapping occurs between distinct predicted masks on the same frame, and (ii) the predicted masks extend beyond the boundaries of the corresponding gtbox. To address these concerns, we introduce a pseudo mask optimization strategy, DOOB (Deleting the Overlapping and Out-of-Boundary sections). Specifically, DOOB involves the removal of overlapping portions of pseudo masks from different instances and the removal of parts extending beyond the gtbox boundaries.

With our multi-keyframes tracking strategy, we obtain a combination of high-quality pseudo masks, named Track-masks-final, selected from the three aforementioned sets (HQ-SAM-masks, IDOL-BoxInst-masks, Track-masks) using SHQM (Selection of High-Quality Masks). Specifically, the SHQM method determines the higher-quality pseudo mask among the three by summing the SCM scores calculated between the target pseudo-mask and the other two pseudo-masks, where a higher SCM score indicates higher quality. As illustrated in Fig. 1 (OURS), Track-masks-final exhibits significantly improved visual quality compared to the previous pseudo masks, closely resembling the gtmasks.

Fig. 4 illustrates the workflow of high-performance box-supervised VIS, involving a two-step strategy. In Step 1, we obtain a training set with high-quality pseudo masks by leveraging three segmentation models. In Step 2, the PM-VIS model is trained using a combination of box annotations and the aforementioned pseudo masks. Through these procedures, our PM-VIS model achieves SOTA performance.

The pipeline in Fig. 5 summarizes our method concisely for generating high-quality pseudo masks through two stages. In Stage 1, we employ IDOL-BoxInst and HQ-SAM to process the training dataset under box supervision, yielding two types of pseudo masks: IDOL-BoxInst-masks and HQ-SAM-masks. We establish correspondence between IDOL-BoxInst-masks and gtboxes by calculating SCM scores with the assistance of HQ-SAM-masks and gtboxes. Additionally, the proposed post-processing optimization strategy, DOOB, is applied to both types of pseudo masks obtained in Stage 1, resulting in refined pseudo masks. In Stage 2, we employ VOS model DeAOT to track instances in the video and obtain high-quality pseudo masks. At this stage, there are two points that we need to note specifically. Firstly, we initialize the model using pseudo masks from keyframes and subsequently track the pixel-level prediction information for the targets throughout the entire video. Among these, keyframes are selected from the positions of certain instance pseudo masks in IDOL-BoxInst-masks determined by the SCM scores calculated between IDOL-BoxInst-masks, HQ-SAM-masks, and gtboxes. Secondly, to further enhance pseudo mask quality, we calculate the confidence scores (SHQM) for HQ-SAM-masks, IDOL-BoxInst-masks, and Track-masks. Then, based on these scores, we choose the higher-quality pseudo masks from HQ-SAM-masks or IDOL-BoxInst-masks to substitute for the instance pseudo masks in the original Track-masks.

The Projection Loss [12] utilizes gtbox annotation to supervise the horizontal and vertical projections of the predicted mask, enhancing the alignment between the predicted mask and the gtbox region. The Projection Loss is formulated as:

where $L_{dice}$ represents the Dice loss defined in BoxInst [12], $p$ denotes either the horizontal or vertical axis, $\max$ is the max operations along with each axis, $m$ corresponds to the predicted mask, and $b$ signifies the ground-truth box mask.

The Pairwise Affinity Loss [12] supervises pixel-level mask predictions without pixel-wise annotations by considering the color similarity between pixels. Specifically, for two pixels at coordinates $(i,j)$ and $(l,k)$, their color similarity is denoted as $S_{e}=\exp\left(-\frac{\parallel c_{i,j}-c_{l,k}\parallel}{\theta}\right)$. Here, $c_{i,j}$ and $c_{l,k}$ represent color vectors in the LAB color space, with $\theta$ as a hyper-parameter (default as 2). The color similarity threshold $\tau$ is set at 0.2, labeling the edges with similarity exceeding this threshold as 1. The network predicts $m_{i,j}\in(0,1)$ as the probability of pixel $(i,j)$ being part of the foreground. Consequently, the probability $y_{e}=1$ for pairwise mask affinity is formulated as $P(y_{e}=1)=m_{i,j}\cdot m_{l,k}+(1-m_{i,j})\cdot(1-m_{l,k})$. The Pairwise Affinity Loss is defined as:

where $E_{in}$ is the set of edges containing at least one pixel in the box, and $N$ is the number of edges in $E_{in}$. The indicator function $\mathbbm{1}_{\{S_{e}>\tau\}}=1$ if $S_{e}>\tau$, and $0$ otherwise.

By substituting the gtmasks with the pseudo masks, the pixel-wise mask supervision losses in the pixel-supervised VIS methods can be extended to the pseudo masks supervision losses, comprising the Focal Loss [17] and the Dice Loss [16].

The IDOL-BoxInst algorithm is a box-supervised VIS method that combines the principles of box-supervised losses [12] with the IDOL algorithm [2], effectively eliminating the need for mask annotations and related losses [16, 17]. PM-VIS is trained using a combination of spatio-temporal losses, including BoxInstLoss [12], which comprises Projection Loss and Pairwise Affinity Loss from BoxInst [12], as well as pixel-wise mask supervision losses (Focal loss [17] and Dice loss [16]), referred to as MaskLoss. PM-VIS adopts a weight configuration similar to that of IDOL and BoxInst, setting the weight (W1) for BoxInstLoss to 1 and the weight (W2) for MaskLoss to 0.5.

The HQ-SAM-masks, obtained by processing video frames based on the proposals of gtboxes using the HQ-SAM [15] model, serves as pseudo mask annotations for the training set. In this work, we train the PM-VIS model using pseudo-mask supervision on the mentioned datasets. We observe that the VIS performance is better with the use of pseudo masks compared to without using them. However, due to the limited quality of HQ-SAM-masks, the improvement is constrained.

As illustrated in Fig. 5 (Stage 1), we generate another type of pseudo masks, namely IDOL-BoxInst-masks, in addition to the pseudo masks generated by the HQ-SAM model. The IDOL-BoxInst model, trained without mask annotations, generates these pseudo masks by predicting the targets in the training set. Nonetheless, the current pseudo masks obtained from this process exhibit several concerns: (i) the predicted masks might exhibit suboptimal quality and inherent errors, (ii) the labels corresponding to the predicted or tracked masks could be inaccurate, and (iii) multiple distinct masks may be independently predicted for the same target, resulting in pixel-level ambiguity and potentially compromising experimental precision. To address these challenges, we introduce an innovative predictive mask allocation and correction strategy called SCM (as described in Section III-D1) that aims at using HQ-SAM-masks to aid in establishing a coherent relationship between the predicted masks and gtboxes for certain instances. Additionally, we conduct an annotation quantity analysis and find that the number of instance annotations for IDOL-BoxInst-masks is relatively lower compared to the ground-truth datasets. Specifically, as shown in Fig. 6, the quantity of IDOL-BoxInst-masks obtained in our experiments is found to be 1.8%, 3.5%, and 7.3% lower than the number of instances annotated in the ground-truth for YTVIS2019, YTVIS2021, and OVIS, respectively.

As shown in Fig. 5 (Stage 2), Track-masks is a collection of pseudo masks obtained by initializing the semi-supervised VOS model DeAOT with instance mask annotations from IDOL-BoxInst-masks and tracking the instances throughout the entire video. Typically, semi-supervised VOS models require information about the target in the first frame of the video during prediction. However, for our task, the initial frame instance pseudo mask may not be optimal. To overcome this limitation, we propose a mechanism to initiate tracking from various positions in the video, progressing both forwards and backwards until reaching the start and end frames, respectively. Subsequently, the tracked results are amalgamated, and these starting positions are termed keyframes. Specifically, our procedure is as follows: firstly, we employ the SCM method to create a collection of high-quality keyframes from the IDOL-BoxInst-masks. Secondly, we obtain the pseudo mask collection, Track-masks, by feeding the annotations from these keyframes into the DeAOT model for prediction. Finally, we employ two methods to further enhance the quality of Track-masks. These methods involve expanding the keyframes for each instance from one to $len(video)/K$ and selecting high-quality pseudo masks for optimizing Track-masks from both IDOL-Boxlnst-masks and HQ-SAM-masks using the SHQM method. Here, $len(video)$ denotes the video length, and we configure $K$ as a hyper-parameter, setting it to be 10 for YTVIS2019/YTVIS2021 and 100 for OVIS.

We present a methodology to quantify the correlation between pseudo masks, as shown in Eq. (4). Let $m_{1}$ and $m_{2}$ denote two distinct types of pseudo masks. We calculate the IoU between these masks and also compute the IoU [37] between the bounding boxes of these masks and the gtbox. The combination of these three factors generates a score, denoted as the Score of Corresponding Mask (SCM), signifying the relationship between the two pseudo masks.

where $Box(m)$ represents the bounding box of a given pseudo mask, $IOU(\cdot,\cdot)$ represents the calculation of IoU, while $gtbox$ denotes the ground-truth bounding box of the target.

We introduce another technique called SHQM for Selecting the High-Quality Masks among the three variants: HQ-SAM-masks, IDOL-BoxInst-masks, and Track-masks. This method establishes a composite connection between each pseudo mask and the gtbox, generating confidence scores $S_{ij}$ for every pseudo mask. The pseudo mask attaining the highest score, labeled as $m_{\text{z}}$, indicates the superior-quality pseudo mask.

This relationship is established through the assessment of the SCM denoted as $S_{ij}$ between different pairs of pseudo masks $m_{i}$ and $m_{j}$, where $i$ and $j$ are members of the set $\{1,2,3\}$ and are distinct from each other. $S_{ij}$ is equal to $S_{ji}$ here.

After analyzing the pseudo mask datasets, two significant issues in the predicted masks have become evident: (i) overlapping occurrences between distinct predicted masks within the same frame, and (ii) the expansion of predicted masks beyond the boundaries of the corresponding gtbox, as illustrated in Fig. 7. To address these concerns and enhance the accuracy of the pseudo mask training dataset, we introduce an optimization approach named DOOB, which involves Deleting the Overlapping and Out-of-Boundary sections. Specifically, the strategy ascertains ownership of overlapping regions by computing the bounding rectangle of the overlapping mask and gauging the IoU [37] value with the associated gtbox. Subsequently, it eliminates mask annotations for the current target that exceed a reasonable range determined by the gtbox.

During the experiments, it becomes apparent that the number of Track-masks is lower than that of the ground-truth data. Upon visualization, as depicted in Fig. 8, we identify several issues with these missing data, including indistinct instance features, low discriminability of instances from surrounding objects or the environment, and severe occlusions. Thus, we propose a filtering method named Missing-Data to remove unnecessary Missing Data from the ground-truth. Specifically, this method leverages our proposed pseudo mask set, Track-masks-final, as a reference to eliminate corresponding discrepancies in the ground-truth data, resulting in a ground-truth dataset of the same size as Track-masks-final.

Our pseudo masks have substantially improved the performance of weakly supervised VIS. However, it’s inevitable that the pseudo mask data we obtained, Track-masks-final, still exhibit disparities in instance segmentation quality compared to the ground-truth, as shown in Fig. 9. Specifically, Track-masks-final tends to struggle with small, complex, and less distinct instances. While this issue poses a challenge in weakly supervised settings, it can be reframed as an optimization problem when considered under fully supervised conditions. Therefore, we propose a method called RIA. This method involves calculating the mask IoU between gtmasks and Track-masks-final data to map and Remove Instance Annotations in the ground-truth with relatively low mask IoU values. As illustrated in Fig. 9 (gtmasks), datasets characterized by lower IoU values frequently encounter challenges, including inconspicuous features and severe occlusions. Consequently, this refinement and optimization process enhances the quality of the ground-truth data, yielding a more compact yet higher-quality dataset.

YTVIS2019, a subset of YTVOS18/19, comprises 2883 videos, 40 categories, 4483 objects, and over 130k annotations. YTVIS2021, an improved version of YTVIS2019, has expanded its dataset, refined the 40-category label set, and thus partially overlaps with YTVOS18/19. OVIS, one of the most challenging datasets in the VIS domain, presents significant hurdles. Instances within OVIS exhibit severe occlusion, intricate motion patterns, and rapid deformations. The dataset consists of 607 training videos and 140 validation videos, with notably extended video durations averaging 12.77 seconds. Regarding VIS evaluation, we employ standard metrics such as AP, AP50 (Average Precision at 50% IoU), AP75 (Average Precision at 75% IoU), AR1 (Average Recall at 1), and AR10 (Average Recall at 10).

YTVOS18/19 and YTVIS2019/2021 represent two distinct datasets utilized for VOS and VIS tasks, respectively. However, in terms of data sources, all data from YTVIS2019 and some data from YTVIS2021 belong to YTVOS18/19. In this paper, we use DeAOT as a tracking model to obtain pseudo masks. To ensure the credibility of our task and avoid the impact of overlapping data, we exclude the overlapping data when retraining DeAOT using data from YTVOS18/19, as shown in Table II. The overlapping data between YTVIS2019/2021 and YTVOS18/19 accounts for 64.5% of the data in YTVOS18/19, while the non-overlapping data accounts for 35.5%. As for VOS evaluation, we adopt the evaluation approach from the DeAOT algorithm, including metrics like the region similarity $J$, the contour accuracy $F$, and their mean value (referred to as $J\&F$).

Although the pseudo masks we use provide a representation of the target within a specific range, they still exhibit limitations in accurately delineating the target. Issues such as significant pixel color variations across different parts of the target and color similarities among distinct targets pose challenges to the precision of target mask annotations. To address this problem, we introduce BoxInstLoss, a technique that leverages inter-pixel information. On one hand, this approach compensates for the incompleteness in capturing target boundaries. On the other hand, it addresses imprecision in target mask annotation by capitalizing on color similarities between pixels. In Table III, we investigate the impact of the weight W1 of BoxInstLoss on VIS AP when training the model with pseudo masks. The best result is achieved at W1 = 1, while setting W1 = 0 means that the algorithm does not utilize BoxInstLoss. Our algorithm achieves a 2.1% increase in AP, reaching 48.7%, when using BoxInstLoss.

Table IV demonstrates the influence of the weight (W2) assigned to MaskLoss on the algorithm. The optimal result is achieved with W2 = 0.5, whereas setting W2 = 0 means excluding MaskLoss and pseudo masks. The results indicate a noteworthy enhancement in performance with the incorporation of pseudo masks, resulting in a 3.9% increase in AP, elevating it to 48.7%. While BoxInstLoss, built upon bounding boxes, partly simulates real mask effects, this weakly supervised approach lacks object shape information and can be susceptible to object color variations. Consequently, its performance distinctly deviates from ground-truth conditions. Certainly, our proposed pseudo masks, by providing a rough contour of the object, enhance the algorithm’s discernment of object characteristics such as shape, position, and size, thereby contributing to the algorithm’s overall performance.

Table V presents a comparison of various pseudo masks under the PM-VIS algorithm. It is evident from the AP performances of the three pseudo masks variants in the PM-VIS algorithm that Track-masks-final exhibits superior quality, achieving 48.7% AP, while HQ-SAM-masks performs less favorably, attaining 46.8% AP. For HQ-SAM-masks and IDOL-BoxInst-masks, we effectively address the issues of overlap and out-of-bounds in the initial pseudo masks using the DOOB strategy. When applied to IDOL-BoxInst-masks, this strategy improves the AP performance by 1.2%, reaching 46.9%. When considering the highest-quality pseudo masks, Track-masks-final, the application of the SHQM method for single-keyframe tracking does not result in an improved algorithm performance. In fact, it leads to a 0.8% decrease in AP. This outcome can be attributed to the predominance of IDOL-BoxInst-masks and HQ-SAM-masks in terms of quantity. The SHQM method, when selecting pseudo masks, appears to be misled by these lower-quality pseudo masks, inadvertently introducing a surplus of unnecessary errors. Conversely, in the context of multi-keyframe tracking, the resultant pseudo masks consist of multiple sets of relatively high-quality masks. When selecting pseudo masks using the SHQM method in this scenario, it introduces a significant number of better-quality tracking masks. This leads to a 0.5% relative increase in AP compared to single-frame scenarios, ultimately reaching an optimal level of 48.7% AP.

Among the two types of pseudo masks with an equal number of instance annotations, IDOL-BoxInst-masks and Track-masks-final, the annotation quality of IDOL-BoxInst-masks is observed to be lower than that of Track-masks-final, as evident from the results of weak supervision. In some cases, there are instances that are expected to be recognized or segmented more effectively by weak supervision methods, but the results in IDOL-BoxInst-masks are less satisfactory. On the other hand, Track-masks-final recognizes fewer of these instances. As shown in Table VI, we conduct a comparative analysis of the results obtained when using these two types of pseudo masks as aids in filtering ground-truth data. The results show that training with PM-VIS on ground-truth data filtered with the assistance of Track-masks-final outperforms IDOL-BoxInst-masks by 1.3%, achieving an AP of 50.0%.

Table VII presents the performance of our PM-VIS algorithm compared to the IDOL algorithm without BoxInstLoss on ground-truth data. The training set used here comprises YTVIS2019 ground-truth data that has undergone two filtering methods: Missing-Data and RIA. These results showcase the algorithm performance on YTVIS2019 VAL. The introduction of BoxInstLoss in our algorithm is primarily motivated by its ability to examine instance segmentation from a pixel similarity perspective. Therefore, we believe that it can still be effective in a fully supervised scenario. As observed, with the inclusion of BoxInstLoss, our algorithm demonstrates a notable 2.3% increase in AP, reaching 50.0%.

Table VIII illustrates the impact of two filtering and optimization methods, Missing-Data and RIA, on algorithm performance. It’s evident that the current algorithms struggle to handle the issues present in the filtered-out data. When training the IDOL-BoxInst model using all ground-truth data without any pixel-level annotations, we notice that the algorithm AP decreases by 0.9% compared to when the missing data is not used. Training the PM-VIS model on ground-truth data filtered only by Missing-Data leads to a slight improvement in AP, with a 0.2% increase compared to using all ground-truth data. Furthermore, when training the PM-VIS model on ground-truth data filtered using both methods, the trained model achieves a significant increase in AP, improving by 0.9% to reach 50.0%.

Table IX presents the impact of different mask IoU thresholds on RIA regarding algorithm performance. These thresholds are determined by calculating mask IoU between Track-masks-final and ground-truth data. When $\tau$ is set to 0, it indicates that this filtering method is not applied, whereas $\tau$ greater than 0 signifies the removal of instance annotations from ground-truth data when their mask IoU values fall below $\tau$. It is worth to note that the mask IoU is computed between these pseudo masks and ground-truth data because lower mask IoU values correspond to lower mask quality. Despite employing various fusion strategies, enhancing the quality of these instance pseudo masks proves challenging. Consequently, we can conclude that these instances are difficult to identify in certain video frames. Removing them has the potential to improve algorithm performance. In our experiments, setting $\tau$ to 0.6 leads to a 0.7% improvement in the algorithm AP compared to not using this filtering method, achieving a performance level of 50.0%.

As shown in Table X, it is apparent that with the ResNet-50 backbone, our basic IDOL-BoxInst model, trained solely on box annotations, underperforms MaskFreeVIS [9] by a margin of 2.7% AP on YTVIS2019. Upon the incorporation of pseudo masks into the PM-VIS algorithm, we observe a noteworthy 4.8% improvement in AP compared to IDOL-BoxInst. Moreover, this enhancement outperforms the performance of the only box-supervised VIS algorithm, MaskFreeVIS, by 2.1%, achieving an outstanding AP of 48.7%. Similarly, on the YTVIS2021 and OVIS datasets, our IDOL-BoxInst algorithm achieves competitive results, surpassing the MaskFreeVIS and even outperforming some fully supervised VIS algorithms. After integrating pseudo masks into the PM-VIS algorithm, we surpass all box-supervised VIS algorithms, outperforming MaskFreeVIS by 3.7% and 12.1%, thus establishing a new SOTA.

Furthermore, training the PM-VIS model using ground-truth data filtered with pseudo masks also yields improved results for three primary reasons. Firstly, the BoxInstLoss enhances mask segmentation precision by extending constraints to both bounding box ranges and pixel similarity. Secondly, a thorough examination, including visual inspection and model performance analysis, demonstrates that the quality of the acquired pseudo masks (Track-masks-final) is approaching that of gtmasks. Thus, the main factor affecting weakly supervised performance appears to be the presence of low-quality instance mask annotations. The application of our proposed filtering methods (Missing-Data and RIA) retains higher-quality, less ambiguous data. Finally, especially for models employing backbone with limited feature extraction capabilities like ResNet-50, the presence of distinct instance features allows the algorithm to efficiently and accurately acquire necessary information, minimizing the need for extensive feature learning and resulting in faster, more precise utilization of the acquired information. When training the PM-VIS model on the filtered ground-truth data, the AP surpasses IDOL^† on all three datasets, achieving 50.0%, 47.7%, and 29.9% on each, respectively. It’s evident that our collection of pseudo masks not only enhances the performance of box-supervised VIS but also contributes to an improvement in fully supervised VIS to a certain extent.

As presented in Table XI, we also conduct experiments using Swin-L as the backbone on YTVIS2019, YTVIS2021 and OVIS. The comparison reveals that our algorithm PM-VIS outperforms the basic IDOL-BoxInst by 3.2%, 2.6%, and 5.3% on the respective datasets. Similarly, we also surpass MaskFreeVIS by 4.4% on the YTVIS2019 dataset, establishing a new SOTA for box-supervised VIS algorithms. The outstanding performance significantly narrows the performance gap between fully supervised and weakly supervised VIS methods.

Furthermore, in theory, training the PM-VIS model with filtered ground-truth data will also result in performance exceeding IDOL^† with the Swin-L backbone. As our analysis during the training of models with ResNet-50 suggests, stronger backbones, such as Swin-L, possess a greater ability to extract information from ambiguous or low-quality instance segmentation annotations. When utilizing the Swin-L backbone, the model will extract target information from the data more precisely. This implies that by using less data, specifically excluding non-essential data from the dataset, the model can still achieve or even surpass the performance obtained by using the entire dataset. We train the PM-VIS model using Swin-L backbone on the same filtered ground-truth data used for training with ResNet-50 backbone, and the model consistently delivers the expected performance. Specifically, on the YTVIS2019, YTVIS2021, and OVIS datasets, the fully supervised VIS model, PM-VIS, outperforms IDOL^† in terms of AP, achieving 63.0%, 59.2%, and 41.0%, respectively. This illustrates that our pseudo masks enhance the performance of weakly supervised algorithms and lead to improvements in fully supervised scenarios, regardless of whether a stronger or weaker backbone is employed.