DaTaSeg: Taming a Universal Multi-Dataset Multi-Task Segmentation Model

Observing the close relationship among panoptic, semantic and instance segmentation tasks, we propose to train a universal multi-dataset multi-task segmentation model: DaTaSeg.We use a shared representation (mask proposals with class predictions) for all tasks. To tackle task discrepancy, we adopt different merge operations and post-processing for different tasks. We also leverage weak-supervision, allowing our segmentation model to benefit from cheaper bounding box annotations. To share knowledge across datasets, we use text embeddings from the same semantic embedding space as classifiers and share all network parameters among datasets. We train DaTaSeg on ADE semantic, COCO panoptic, and Objects365 detection datasets. DaTaSeg improves performance on all datasets, especially small-scale datasets, achieving 54.0 mIoU on ADE semantic and 53.5 PQ on COCO panoptic. DaTaSeg also enables weakly-supervised knowledge transfer on ADE panoptic and Objects365 instance segmentation. Experiments show DaTaSeg scales with the number of training datasets and enables open-vocabulary segmentation through direct transfer. In addition, we annotate an Objects365 instance segmentation set of 1,000 images and will release it as a public benchmark.

Paper

References (79)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 8rwi5/10 · confidence 5/52023-07-03

Summary

This paper proposes the universal model for image segmentation using multi-dataset multi-task training. By using universal segmentation representations at the entity (thing or stuff) level, the paper use merge operation for different segmentation task. Experiments show the effectiveness of the proposed method.

Strengths

1, The paper is well-written and easy to understand. 2, The extensive experiments show the effectiveness of multi-dataset multi-task training.

Weaknesses

1, Limited novelties. The main contribution of this paper is multi-dataset multi-task training. However, there is no obvious difference to OneFormer except for the multi-dataset training. To solve the label space conflict of multi-dataset training, the paper directly uses the language embeddings of the language model that has been widely explored. 2, Merging universal segmentation representations seems like following the fine-to-coarse pipeline. There is no obvious clue to prove whether this fine-to-coarse is better than the coarse-to-fine pipeline. 3, Weakly-supervised instance segmentation module also has limited novelties by just using the projection loss of BoxInst to add the larger dataset objects365.

Questions

1, Could the author try the other sampling strategy in multi-dataset training? In the paper, it would lead to the imbalance of dataset sampling. 2, The performance improvement on COCO mainly comes from using the objects365 dataset. Direct pretraining object365 and then finetuning to COCO could perform better than the proposed method in the paper.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

This paper cannot solve the annotation inconsistency of multi-dataset training. For example, the `window' defined in COCO and ADE20K are different where COCO stuff but ADE20K thing.

Reviewer LhuR6/10 · confidence 4/52023-07-05

Summary

This paper proposes DaTaSeg, a universal multi-dataset multi-task segmentation model. DaTaSeg uses a shared representation and different merge operations and post-processing for different tasks. Weak-supervision is employed for cheaper bounding box annotations and knowledge is sharing across different datasets with text embeddings from the same semantic embedding space and shared network parameters. A subset of the Objects365 validation set is annotated for instance segmentation. Experiments shows that DaTaSeg gets improved performance on dataset-specific models and enables weakly-supervised knowledge transfer. DaTaSeg also scales with the training dataset number and enables open-vocabulary segmentation.

Strengths

1) This paper is easy to follow. 2) The method proposed in this paper will be useful in the future as it achieves the segmentation task under the multi-task multi-dataset setting.

Weaknesses

The ablation study in the paper is somewhat not sufficient. The hyperparameters λ in equation (4) and μ in equation (5) are not provided and there needs to be experiments on the impact of hyperparameters.

Questions

This paper proposes the multi-dataset multi-task segmentation model. The hyperparameters λ and μ in the equations (4) and (5) are crucial for the training. Could you provide the experiments on the impact of hyperparameters on the model performance?The authors do not discuss Limitations in the paper.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors do not discuss Limitations in the paper.

Reviewer t7xz4/10 · confidence 5/52023-07-06

Summary

[Tasks] This paper introduces DaTaSeg, a universal multi-dataset multi-task segmentation model that addresses the interconnections between panoptic, semantic, and instance segmentation tasks. [Methods] DaTaSeg utilizes a shared representation, consisting of mask proposals with class predictions, across all tasks. To handle task discrepancies, the model employs distinct merge operations and post-processing techniques tailored to each task. Additionally, DaTaSeg leverages weak supervision, enabling cost-effective bounding box annotations to enhance the segmentation model. To facilitate knowledge sharing across datasets, DaTaSeg utilizes text embeddings from the same semantic embedding space as classifiers and shares all network parameters among datasets. [Experiments] The model is trained on ADE semantic, COCO panoptic, and Objects365 detection datasets. DaTaSeg exhibits improved performance across all datasets, particularly for small-scale datasets, achieving a 54.0 mIoU on ADE semantic and a 53.5 PQ on COCO panoptic. Furthermore, DaTaSeg enables weakly-supervised knowledge transfer for ADE panoptic and Objects365 instance segmentation. [Results] Experimental results indicate that DaTaSeg scales effectively with the number of training datasets and facilitates open-vocabulary segmentation through direct transfer. [Dataset] Additionally, the authors have annotated an Objects365 instance segmentation set comprising 1,000 images, which will be released as a public benchmark.

Strengths

1. The authors have conducted comprehensive experiments, showcasing the state-of-the-art performance achieved on multiple long-tailed recognition benchmarks. This highlights the robustness and effectiveness of their proposed method. 2. The paper is skillfully organized, making it easy to follow. The logical structure and clear presentation enhance the reader's understanding of the research. 3. Efficient Use of Supervision. While previous works have explored training universal models on multiple datasets and tasks, a notable strength of this paper is the effective utilization of weak bounding box supervision for segmentation. Compared to full mask annotations, weak bounding box supervision is a more cost-effective and practical solution. This approach makes the proposed method more accessible and applicable in real-world scenarios.

Weaknesses

[Technical contributions on multi-task multi-task model.] One of the key contributions of this paper is the joint training of multiple datasets and multiple tasks within a unified framework. However, it should be noted that this aspect has already been explored by [1]. The referenced work employs a single set of parameters pre-trained for Semantic/Instance/Panoptic Segmentation, Referring Segmentation, Image Captioning, and Image-Text Retrieval tasks. Therefore, it might be necessary for the authors to clarify the differences between this work and X-Decoder. [Technical contributions on text embedding classifier.] The utilization of a text embedding classifier has also been explored in previous works, such as [2]. In this work, an image encoder is trained to encode pixel embeddings, while CLIP text embeddings are employed as the per-pixel classifier. The key ideas are quite similar, although there are some differences on how to better leverage the text embeddings. [Lack of comparisons with some published works and subpar performance compared to state-of-the-art methods.] The paper primarily compares the model's performance against methods designed or trained solely on a single task, referred to as "Separate" in the paper. However, it is worth noting that DaTaSeg benefits from training on a significantly larger sample size compared to these listed works. Consequently, the observed performance gains over the "Separate" models are not unexpected. Furthermore, there are published works, such as X-Decoder [1], that jointly train models on multiple datasets and tasks, yielding superior performance on many benchmarks compared to the reported results in this paper. On ADE semantic (mIoU): X-Decoder achieves 58.1 compared to DaTaSeg's 54.0; on COCO Panoptic (PQ): X-Decoder outperforms DaTaSeg with 56.9 versus 53.5. In most benchmarks, DaTaSeg performs worse than X-Decoder. [1] "Generalized decoding for pixel, image, and language." Zou, Xueyan, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai et al. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 15116-15127. 2023. [2] "Language-driven Semantic Segmentation." Boyi Li and Kilian Q Weinberger and Serge Belongie and Vladlen Koltun and Rene Ranftl. International Conference on Learning Representations. 2022

Questions

Most of my questions are listed in the Weaknesses section, with my main concerns focused on the technical contributions and performance comparisons with some published works. However, I have a couple of minor questions: 1. Will the code and model weights be made available to the public? It would be valuable to have access to these resources for further exploration and replication of the proposed method. 2. Given that this model can be trained with weakly-supervised tasks, I'm curious if further improvements in performance can be achieved by training on a larger quantity of weakly-supervised samples. Are there any trade-offs between the quality and quantity of the samples when it comes to enhancing model performance? I believe addressing these questions would provide additional insights into the practicality and potential improvements of the proposed method.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes, the authors adequately addressed the limitations

Reviewer 7Lho7/10 · confidence 4/52023-07-07

Summary

This paper proposes DaTaSeg, a general multi-dataset multi-task segmentation model. It utilizes shared representations and different pooling operations to perform panoramic, semantic and instance segmentation tasks. DaTaSeg benefits from weak supervision and knowledge transfer across datasets. It outperforms separate training on all datasets (especially smaller ones) and enables weakly supervised segmentation. The model also transfers well to unseen datasets and supports open vocabulary segmentation.

Strengths

The paper addresses the challenge of training a single model on multiple segmentation tasks and datasets by proposing DaTaSeg, which has the potential to save computational resources and streamline the development of segmentation models. It can also benefit from weak supervision by incorporating cheaper bounding box annotations. The proposed DaTaSeg shows promising results in transferring to unseen datasets and enabling open-vocabulary segmentation. The authors annotate a subset of the Objects365 dataset and plan to release it as a public benchmark for instance segmentation. This contributes to the research community by providing a standardized evaluation dataset.

Weaknesses

The Segment Anything model (SAM) released in April 2023 should be included for comparison as it's also a universal segmentation model which has zero/few-shot capability.

Questions

May need proofreading: Line 35, "which are" -> "which is" Line 38, "which map" -> "which maps"

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

According to the F section in Supplementary Material, the computational cost is quite high which may affect the reproducibility. Consider releasing the pre-trained models in various sizes.

Reviewer EhPn6/10 · confidence 5/52023-07-08

Summary

This paper introduces DaTaSeg, a universal multi-dataset multi-task segmentation model. It uses a shared representation for panoptic, semantic, and instance segmentation tasks, with different techniques to address task differences. Weak supervision and knowledge sharing are employed. DaTaSeg improves performance on various datasets, especially small-scale ones. It enables knowledge transfer and open-vocabulary segmentation. An Objects365 instance segmentation set of 1,000 images will be released as a benchmark.

Strengths

1. It is noticed that training segmentation models on multiple datasets to obtain better results is difficult. It is more challenging to train segmentation models on multiple datasets for multiple tasks (instance segmentation, semantic segmentation, weakly supervised instance segmentation. The performance of DaTaSeg on COCO and ADE shows the effectiveness of the proposed approach 2. The paper leverages box-level supervision to improve the segmentation performance.

Weaknesses

1. [minor] The definition of mask proposal is commonly used in instance segmentation [1]. However, the description uses too many math terms to describe a simple concept that can be demonstrated by natural language, figures, and more compact math. This is not good for readers to understand. 2. [major] The results of weakly supervised instance segmentation look not promising. In the previous state-of-the-art method, weakly supervised instance segmentation methods achieves 70-90% performance (mAP) as what their fully supervised versions do [2,3]. However, in the 2, the instance segmentation of Object365 looks not promising, only 10ish%. What is the upper-bound performance of O365 instance segmentation? 3. [major] The approach introduces the unified mask representation without comparison or discussion with other mask representations. It is well-known that training multi-dataset or multi-tasks is not trivial, and it is better to show more training details and empirical design considerations in the paper. [1] run-length encoding: https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocotools/coco.py#L265 [2] Lan, Shiyi, et al. "Vision transformers are good mask auto-labelers." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. [3] Li, Wentong, et al. "Box2Mask: Box-supervised Instance Segmentation via Level-set Evolution." arXiv preprint arXiv:2212.01579 (2022).

Questions

None

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

None

Reviewer t7xz2023-08-15

I would like to extend my appreciation to the authors for their efforts in addressing my questions. While some of my concerns have indeed been adequately resolved, I must express that I still have reservations regarding the technical contributions and the comparisons made with certain published works. 1. The authors have argued that comparing the X-Decoder (XD) should be avoided due to its recent publication at CVPR 2023. I acknowledge this perspective, however, it's worth noting that ***X-Decoder was made available on ArXiv about 9 months*** ago and was accepted at CVPR around 5 months ago. Considering the dynamic nature of this field, I believe it is not unreasonable to consider a comparison. 2. In relation to the task variation, the authors have contended that XD's advantage largely stems from being trained on more data. It's important to note that the additional data largely comes from ***image-text paired data***, which ***is potentially easier to label than detection datasets***. I regard the ability to encompass a broader set of tasks as a strength rather than a weakness, particularly considering that the performance gains and the capability of supporting more tasks are attained through training on relatively *cheap* image-text pairs. \ Additionally, I would like to bring to the authors' attention the existence of OpenSeeD (accepted at ICCV2023) [2], which also harnesses detection data to enhance segmentation tasks. *Please note that I do not expect a comparison to [2] to be drawn. It's merely offered for your reference.* 3. The authors point out that the primary contributions stem from "include a simple weakly-supervised instance segmentation training using bounding-box supervision to increase the training data". Nonetheless, it's worth noting that weakly supervised instance segmentation is already a well-established task, and ***the Projection Loss employed in the paper to facilitate this approach is directly borrowed from [1]***. Given these factors, I find it challenging to regard this as a strong technical contribution. Furthermore, it can be argued which of the two is actually less resource-intensive: detection data or image-text data. Thank you once again for your time on addressing my questions! [1] Tian, Z., Shen, C., Wang, X. and Chen, H., 2021. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 5443-5452). [2] Zhang, H., Li, F., Zou, X., Liu, S., Li, C., Gao, J., Yang, J. and Zhang, L., 2023. A simple framework for open-vocabulary segmentation and detection. arXiv preprint arXiv:2303.08131.

Authorsrebuttal2023-08-19

Dear Reviewer t7xz, We thank the reviewer for their review efforts and additional comments. Before we address the concerns, we would like to emphasize that as mentioned in the rebuttal, X-Decoder is a pioneering work and we are very willing to cite and compare it in our revision. Now, we carefully address the concerns below. >***Comparison with X-Decoder***: We simply added a note that X-Decoder is published at CVPR 2023. We did not argue that comparing with X-Decoder should be avoided --- We promised in our rebuttal W1: *"we'll cite X-Decoder and include the comparison in the revision"*. Given the several key differences (rebuttal W1 and W3) between our work and X-Decoder, it's still hard to have a strict apple-to-apple comparison though. Below, we highlight the differences again: 1. **Task and Focus:** The pre-training in X-Decoder includes panoptic segmentation, referring segmentation, and image-text pairs (image-text retrieval and image captioning), which has a focus on vision-language tasks. These tasks are very different from ours: we cotrain on three mainstream segmentation tasks (panoptic/semantic/instance) and exclusively focus on segmentation. 2. **Training paradigm:** We directly cotrain on multiple datasets using a shared set of parameters (single model), while the ADE and COCO results in X-Decoder Table 1 is **task-specific transfer**. That is, X-Decoder first pretrains on large-scale data and then fine-tunes on each target dataset using different sets of fine-tuned parameters. 3. **Training data:** X-Decoder pretrains on COCO panoptic, referring segmentation and image-text pairs with a long schedule, and we quote from their paper: *"all the pre-trained models are trained with 50 epochs of COCO data and roughly 45 epochs of 10 million image-text pairs"*. Afterwards, they fine-tune the model on COCO and ADE20K separately. By contrast, we co-train on COCO panoptic, ADE20K semantic, and weakly-supervised Objects365 v2 detection with a total of 1.8M of training images. We do not train on large amounts of text data. 4. **Model architecture:** X-Decoder adopts Mask2Former. Our model architecture is different as explained in supplementary Sec. E. More importantly, we don't have an ***online text encoder*** in our model architecture. Consequently, having a strict apple-to-apple comparison will require a careful alignment of each setting, which is beyond the scope of this work and rebuttal. >***Image-text vs. detection data:*** We agree with the reviewer that *the ability to encompass a broader set of tasks as a strength rather than a weakness*, and we never claim that it is a weakness. This is also one of our motivations to include various types of segmentation annotations from multiple datasets for training a single model. We emphasize that both image-text paired data and detection data are valuable training data. However, it is beyond the scope of this work and rebuttal to compare which one is better. Finally, we thank the reviewer for bringing OpenSeeD to our attention. We are also happy to cite OpenSeeD in our final revision. >***Projection loss:*** We thank the reviewer for the comment. We emphasize that we never claim that it is our *primary contribution* to use the projection loss. We have carefully phrased our contributions in the draft and rebuttal. Our primary novelty is that we **propose a single unified framework for multi-dataset multi-task segmentation**. We list other contributions below: 1. With our multi-dataset multi-task training, we 1) leverage knowledge from multiple datasets, containing different types of annotations, to improve segmentation performance on all datasets, especially smaller-scale datasets, such as ADE20K; 2) enable weakly-supervised instance and panoptic segmentation through knowledge transfer; 3) directly enable open-vocabulary segmentation. 2. We propose to use a universal segmentation representation (Sec. 3.1) with a fully-shared network architecture (Sec. 3.4), and apply different losses to different types of segmentation tasks. With the projection loss, we perform bounding-box-supervised instance segmentation *without any modification to the architecture*, which also performs well in the co-training empirically.

Reviewer 8rwi2023-08-18

Thanks for the author's effort in rebuttal. It still cannot convince me very much. In Table 2, using both COCO panoptic and O365 datasets brings 18.3 mIoU on ADE semantic task. That means the annotation inconsistency still exists to some extent. It is similar to the setting about training the model on ADE semantic and O365 box and inference it on COCO Panoptic. Another reason is the image and task domain gap. That might be why the paper uses the ADE semantic instead of the ADE panoptic dataset. Also, merge operations can somewhat solve the annotation inconsistency. When doing inference, the user should select whether to adopt the merge operation based on different datasets. Anyway, I appreciate the difficulties of multi-dataset training and the heavy workload. Thus, I tend to vote for borderline acceptance. However, I cannot rate higher scores because the problem of the multi-dataset training for segmentation still needs to be solved.

Authorsrebuttal2023-08-19

Dear Reviewer 8rwi, Thank you for recognizing the difficulties in our problem setting and our hard work! We carefully address your comments below. - **Table 2 performance:** When not training on ADE semantic or COCO panoptic, one primary reason for the lower performance is that the model only has limited knowledge about the categories in the untrained dataset. While it's interesting to transfer to untrained datasets, it's not the main problem that our submission is trying to address. - **Why we use ADE semantic instead of ADE panoptic:** 1. In our problem setting, we want to train on a suite of datasets of different segmentation tasks. ADE20K is one of the most widely-used benchmarks for semantic segmentation, so we choose to train on ADE20K semantic dataset. If we also train on ADE20K panoptic dataset, then the performance improvement on ADE20K semantic may come from ADE20K panoptic training (we note that semantic annotation is a subset of panoptic annotation which includes both semantic categories and instance identities), rather than from cotraining on other datasets -- which makes it harder to argue the benefits of multi-dataset multi-task training. 2. Besides, we are also curious about how well our model can perform on ADE20K panoptic without directly training on it, while only exploiting the weaker semantic annotations. - **Merge operation:** During inference time, the type of merge operation applied is decided by the desired segmentation task, as discussed in L151 of the paper. The user only needs to specify the segmentation task to perform. - **The problem of the multi-dataset training for segmentation is not solved:** Yes, we agree. This field is underexplored. Given the great benefits of multi-dataset multi-task segmentation models, we see our paper as one of the very first few works that take an initial step in this direction. We look forward to more future work to further improve the performance.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC