MM1.5: Methods, Analysis & Insights from Multimodal LLM Fine-tuning

We present MM1.5, a new family of multimodal large language models (MLLMs) designed to enhance capabilities in text-rich image understanding, visual referring and grounding, and multi-image reasoning. Building upon the MM1 architecture, MM1.5 adopts a data-centric approach to model training, systematically exploring the impact of diverse data mixtures across the entire model training lifecycle. This includes high-quality OCR data and synthetic captions for continual pre-training, as well as an optimized visual instruction-tuning data mixture for supervised fine-tuning. Our models range from 1B to 30B parameters, encompassing both dense and mixture-of-experts (MoE) variants, and demonstrate that careful data curation and training strategies can yield strong performance even at small scales (1B and 3B). Additionally, we introduce two specialized variants: MM1.5-Video, designed for video understanding, and MM1.5-UI, tailored for mobile UI understanding. Through extensive empirical studies and ablations, we provide detailed insights into the training processes and decisions that inform our final designs, offering valuable guidance for future research in MLLM development.

From Text to Insight: Leveraging Large Language Models for Performance Evaluation in Management

This study explores the potential of Large Language Models (LLMs), specifically GPT-4, to enhance objectivity in organizational task performance evaluations. Through comparative analyses across two studies, including various task performance outputs, we demonstrate that LLMs can serve as a reliable and even superior alternative to human raters in evaluating knowledge-based performance outputs, which are a key contribution of knowledge workers. Our results suggest that GPT ratings are comparable to human ratings but exhibit higher consistency and reliability. Additionally, combined multiple GPT ratings on the same performance output show strong correlations with aggregated human performance ratings, akin to the consensus principle observed in performance evaluation literature. However, we also find that LLMs are prone to contextual biases, such as the halo effect, mirroring human evaluative biases. Our research suggests that while LLMs are capable of extracting meaningful constructs from text-based data, their scope is currently limited to specific forms of performance evaluation. By highlighting both the potential and limitations of LLMs, our study contributes to the discourse on AI role in management studies and sets a foundation for future research to refine AI theoretical and practical applications in management.

SlowFast-LLaVA: A Strong Training-Free Baseline for Video Large Language Models

We propose SlowFast-LLaVA (or SF-LLaVA for short), a training-free video large language model (LLM) that can jointly capture the detailed spatial semantics and long-range temporal context without exceeding the token budget of commonly used LLMs. This is realized by using a two-stream SlowFast design of inputs for Video LLMs to aggregate features from sampled video frames in an effective way. Specifically, the Slow pathway extracts features at a low frame rate while keeping as many spatial details as possible (e.g., with 24x24 tokens), and the Fast pathway operates on a high frame rate but uses a larger spatial pooling stride (e.g., downsampling 6x) to focus on the motion cues. As a result, this design allows us to adequately capture both spatial and temporal features that are beneficial for understanding details along the video. Experimental results show that SF-LLaVA outperforms existing training-free methods on a wide range of video tasks. On some benchmarks, it achieves comparable or even better performance compared to state-of-the-art Video LLMs that are fine-tuned on video datasets.

SkeleTR: Towrads Skeleton-based Action Recognition in the Wild

We present SkeleTR, a new framework for skeleton-based action recognition. In contrast to prior work, which focuses mainly on controlled environments, we target more general scenarios that typically involve a variable number of people and various forms of interaction between people. SkeleTR works with a two-stage paradigm. It first models the intra-person skeleton dynamics for each skeleton sequence with graph convolutions, and then uses stacked Transformer encoders to capture person interactions that are important for action recognition in general scenarios. To mitigate the negative impact of inaccurate skeleton associations, SkeleTR takes relative short skeleton sequences as input and increases the number of sequences. As a unified solution, SkeleTR can be directly applied to multiple skeleton-based action tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It also enables transfer learning and joint training across different action tasks and datasets, which result in performance improvement. When evaluated on various skeleton-based action recognition benchmarks, SkeleTR achieves the state-of-the-art performance.

An In-depth Study of Stochastic Backpropagation

In this paper, we provide an in-depth study of Stochastic Backpropagation (SBP) when training deep neural networks for standard image classification and object detection tasks. During backward propagation, SBP calculates the gradients by only using a subset of feature maps to save the GPU memory and computational cost. We interpret SBP as an efficient way to implement stochastic gradient decent by performing backpropagation dropout, which leads to considerable memory saving and training process speedup, with a minimal impact on the overall model accuracy. We offer some good practices to apply SBP in training image recognition models, which can be adopted in learning a wide range of deep neural networks. Experiments on image classification and object detection show that SBP can save up to 40% of GPU memory with less than 1% accuracy degradation.

MeMOT: Multi-Object Tracking with Memory

We propose an online tracking algorithm that performs the object detection and data association under a common framework, capable of linking objects after a long time span. This is realized by preserving a large spatio-temporal memory to store the identity embeddings of the tracked objects, and by adaptively referencing and aggregating useful information from the memory as needed. Our model, called MeMOT, consists of three main modules that are all Transformer-based: 1) Hypothesis Generation that produce object proposals in the current video frame; 2) Memory Encoding that extracts the core information from the memory for each tracked object; and 3) Memory Decoding that solves the object detection and data association tasks simultaneously for multi-object tracking. When evaluated on widely adopted MOT benchmark datasets, MeMOT observes very competitive performance.

Stochastic Backpropagation: A Memory Efficient Strategy for Training Video Models

We propose a memory efficient method, named Stochastic Backpropagation (SBP), for training deep neural networks on videos. It is based on the finding that gradients from incomplete execution for backpropagation can still effectively train the models with minimal accuracy loss, which attributes to the high redundancy of video. SBP keeps all forward paths but randomly and independently removes the backward paths for each network layer in each training step. It reduces the GPU memory cost by eliminating the need to cache activation values corresponding to the dropped backward paths, whose amount can be controlled by an adjustable keep-ratio. Experiments show that SBP can be applied to a wide range of models for video tasks, leading to up to 80.0% GPU memory saving and 10% training speedup with less than 1% accuracy drop on action recognition and temporal action detection.

Long Short-Term Transformer for Online Action Detection

In this paper, we present Long Short-term TRansformer (LSTR), a new temporal modeling algorithm for online action detection, by employing a long- and short-term memories mechanism that is able to model prolonged sequence data. It consists of an LSTR encoder that is capable of dynamically exploiting coarse-scale historical information from an extensively long time window (e.g., 2048 long-range frames of up to 8 minutes), together with an LSTR decoder that focuses on a short time window (e.g., 32 short-range frames of 8 seconds) to model the fine-scale characterization of the ongoing event. Compared to prior work, LSTR provides an effective and efficient method to model long videos with less heuristic algorithm design. LSTR achieves significantly improved results on standard online action detection benchmarks, THUMOS'14, TVSeries, and HACS Segment, over the existing state-of-the-art approaches. Extensive empirical analysis validates the setup of the long- and short-term memories and the design choices of LSTR.

Semi-TCL: Semi-Supervised Track Contrastive Representation Learning

Online tracking of multiple objects in videos requires strong capacity of modeling and matching object appearances. Previous methods for learning appearance embedding mostly rely on instance-level matching without considering the temporal continuity provided by videos. We design a new instance-to-track matching objective to learn appearance embedding that compares a candidate detection to the embedding of the tracks persisted in the tracker. It enables us to learn not only from videos labeled with complete tracks, but also unlabeled or partially labeled videos. We implement this learning objective in a unified form following the spirit of constrastive loss. Experiments on multiple object tracking datasets demonstrate that our method can effectively learning discriminative appearance embeddings in a semi-supervised fashion and outperform state of the art methods on representative benchmarks.

Stepwise Goal-Driven Networks for Trajectory Prediction

We propose to predict the future trajectories of observed agents (e.g., pedestrians or vehicles) by estimating and using their goals at multiple time scales. We argue that the goal of a moving agent may change over time, and modeling goals continuously provides more accurate and detailed information for future trajectory estimation. In this paper, we present a novel recurrent network for trajectory prediction, called Stepwise Goal-Driven Network (SGNet). Unlike prior work that models only a single, long-term goal, SGNet estimates and uses goals at multiple temporal scales. In particular, the framework incorporates an encoder module that captures historical information, a stepwise goal estimator that predicts successive goals into the future, and a decoder module that predicts future trajectory. We evaluate our model on three first-person traffic datasets (HEV-I, JAAD, and PIE) as well as on two bird's eye view datasets (ETH and UCY), and show that our model outperforms the state-of-the-art methods in terms of both average and final displacement errors on all datasets. Code has been made available at: https://github.com/ChuhuaW/SGNet.pytorch.