As the foundational component of versatile AI applications, training an multimodal large language model (MLLM) relies on multimodal datasets with dynamic modality mixture proportions and sample length distributions. However, existing MLLM systems remain inefficient under dynamic workloads, due to statically coupled decisions of resource allocation and model parallelization between encoders and the LLM backbone. This paper presents MegaScale-Omni, an industrial-grade MLLM training system tailored for dynamic workload adaption and hyper-scale deployment. MegaScale-Omni is built upon the training scheme of encoder-LLM multiplexing with three key innovations: (1) Decoupled parallelism strategies with long-short sequence parallelism for encoders to process variable-length samples, and full-fledged 5D parallelism for the LLM backbone, both organized under a communication-efficient parallelization layout. (2) Unified encoder-LLM representations for flexible, extensible colocation, and a new paradigm of encoder-LLM joint pipeline with workload resilience. (3) Workload balancing techniques via decentralized grouped reordering in data loaders and adaptive resharding from encoder to LLM ranks. MegaScale-Omni is deployed as the foundation of our in-house large-scale MLLM training tasks with thousands of GPUs. Our experimental results demonstrate 1.27×–7.57× throughput improvement under production-grade dynamic workloads, as compared to four state-of-the-art systems.

Weakly supervised instance segmentation with image-level class supervision is a challenging task as it associates the highest-level instances to the lowest-level appearance. Previous approaches for the task utilize classification networks to obtain rough discriminative parts as seed regions and use distance as a metric to cluster pixels of the same instances. Unlike previous approaches, we provide a novel self-supervised joint learning framework as the basic network and consider the clustering problem as calculating the probability that pixels belong to each instance. To this end, we propose our self-supervised joint learning two-stream network (SJLT Net) to finish this task. In the first stream, we leverage a joint learning framework to implement image-level supervised semantic segmentation with self-supervised saliency detection. In the second stream, we propose a Center Detection Network to detect different instances’ centers with the gaussian loss function to cluster instances pixels. Besides, an integration module is utilized to combine information of both streams and get precise pseudo instances labels. Our approach generates pseudo instance segmentation labels of training images, which are used to train a fully supervised model. Our model achieves excellent performance on the PASCAL VOC 2012 dataset, surpassing the best baseline trained with the same labels by 4.6$\%$$AP^r_{50}$on the train set and 2.6$\%$$AP^r_{50}$on the validation set.

For further progress in video object segmentation (VOS), larger, more diverse, and more challenging datasets will be necessary. However, densely labeling every frame with pixel masks does not scale to large datasets. We use a deep convolutional network to automatically create pseudo-labels on a pixel level from much cheaper bounding box annotations and investigate how far such pseudo-labels can carry us for training state-of-the-art VOS approaches. A very encouraging result of our study is that adding a manually annotated mask in only a single video frame for each object is sufficient to generate pseudo-labels which can be used to train a VOS method to reach almost the same performance level as when training with fully segmented videos. We use this workflow to create pixel pseudo-labels for the training set of the challenging tracking dataset TAO, and we manually annotate a subset of the validation set. Together, we obtain the new TAO-VOS benchmark, which we make publicly available at www.vision.rwth-aachen.de/page/taovos. While the performance of state-of-the-art methods on existing datasets starts to saturate, TAO-VOS remains very challenging for current algorithms and reveals their shortcomings.

Weakly supervised localization has attracted increasing attention since only image-wise labels are needed. One mainstream approach, CAM based top-down localization method, suffers from poor resolution and localizing only the most discriminative regions. Another kind, model agnostic perturbation based method, suffers from multiple iterations for each sample. In this paper, we introduce PECA-Net: Preserve and Erase Based Class Attention Networks, which adopts preserve and erase perturbed U-net as the basis, with class activation mechanism as attention to enhance localization capability. Class attention module strengthens informative features and achieves a basic localization. Preserve and erase perturbed U-net replaces the random and iterative extrinsic perturbation with meaningful erasing. In addition, this structure refines the preliminary localization. Since the target object is hit twice, therefore, entitled as double shot. Experiments validate that localization error of both CUB-200 and ILSVRC ImageNet dataset is the new state-of-the-art.