No Arabic abstract
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, THUMOS14, 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.
Spatial and temporal relationships, both short-range and long-range, between objects in videos, are key cues for recognizing actions. It is a challenging problem to model them jointly. In this paper, we first present a new variant of Long Short-Term Memory, namely Relational LSTM, to address the challenge of relation reasoning across space and time between objects. In our Relational LSTM module, we utilize a non-local operation similar in spirit to the recently proposed non-local network to substitute the fully connected operation in the vanilla LSTM. By doing this, our Relational LSTM is capable of capturing long and short-range spatio-temporal relations between objects in videos in a principled way. Then, we propose a two-branch neural architecture consisting of the Relational LSTM module as the non-local branch and a spatio-temporal pooling based local branch. The local branch is utilized for capturing local spatial appearance and/or short-term motion features. The two branches are concatenated to learn video-level features from snippet-level ones which are then used for classification. Experimental results on UCF-101 and HMDB-51 datasets show that our model achieves state-of-the-art results among LSTM-based methods, while obtaining comparable performance with other state-of-the-art methods (which use not directly comparable schema). Further, on the more complex large-scale Charades dataset, we obtain a large 3.2% gain over state-of-the-art methods, verifying the effectiveness of our method in complex understanding.
We propose a method for representing motion information for video classification and retrieval. We improve upon local descriptor based methods that have been among the most popular and successful models for representing videos. The desired local descriptors need to satisfy two requirements: 1) to be representative, 2) to be discriminative. Therefore, they need to occur frequently enough in the videos and to be be able to tell the difference among different types of motions. To generate such local descriptors, the video blocks they are based on must contain just the right amount of motion information. However, current state-of-the-art local descriptor methods use video blocks with a single fixed size, which is insufficient for covering actions with varying speeds. In this paper, we introduce a long-short term motion feature that generates descriptors from video blocks with multiple lengths, thus covering motions with large speed variance. Experimental results show that, albeit simple, our model achieves state-of-the-arts results on several benchmark datasets.
Detecting 3D landmarks on cone-beam computed tomography (CBCT) is crucial to assessing and quantifying the anatomical abnormalities in 3D cephalometric analysis. However, the current methods are time-consuming and suffer from large biases in landmark localization, leading to unreliable diagnosis results. In this work, we propose a novel Structure-Aware Long Short-Term Memory framework (SA-LSTM) for efficient and accurate 3D landmark detection. To reduce the computational burden, SA-LSTM is designed in two stages. It first locates the coarse landmarks via heatmap regression on a down-sampled CBCT volume and then progressively refines landmarks by attentive offset regression using high-resolution cropped patches. To boost accuracy, SA-LSTM captures global-local dependence among the cropping patches via self-attention. Specifically, a graph attention module implicitly encodes the landmarks global structure to rationalize the predicted position. Furthermore, a novel attention-gated module recursively filters irrelevant local features and maintains high-confident local predictions for aggregating the final result. Experiments show that our method significantly outperforms state-of-the-art methods in terms of efficiency and accuracy on an in-house dataset and a public dataset, achieving 1.64 mm and 2.37 mm average errors, respectively, and using only 0.5 seconds for inferring the whole CBCT volume of resolution 768*768*576. Moreover, all predicted landmarks are within 8 mm error, which is vital for acceptable cephalometric analysis.
Most work on temporal action detection is formulated as an offline problem, in which the start and end times of actions are determined after the entire video is fully observed. However, important real-time applications including surveillance and driver assistance systems require identifying actions as soon as each video frame arrives, based only on current and historical observations. In this paper, we propose a novel framework, Temporal Recurrent Network (TRN), to model greater temporal context of a video frame by simultaneously performing online action detection and anticipation of the immediate future. At each moment in time, our approach makes use of both accumulated historical evidence and predicted future information to better recognize the action that is currently occurring, and integrates both of these into a unified end-to-end architecture. We evaluate our approach on two popular online action detection datasets, HDD and TVSeries, as well as another widely used dataset, THUMOS14. The results show that TRN significantly outperforms the state-of-the-art.
Transformers have achieved success in both language and vision domains. However, it is prohibitively expensive to scale them to long sequences such as long documents or high-resolution images, because self-attention mechanism has quadratic time and memory complexities with respect to the input sequence length. In this paper, we propose Long-Short Transformer (Transformer-LS), an efficient self-attention mechanism for modeling long sequences with linear complexity for both language and vision tasks. It aggregates a novel long-range attention with dynamic projection to model distant correlations and a short-term attention to capture fine-grained local correlations. We propose a dual normalization strategy to account for the scale mismatch between the two attention mechanisms. Transformer-LS can be applied to both autoregressive and bidirectional models without additional complexity. Our method outperforms the state-of-the-art models on multiple tasks in language and vision domains, including the Long Range Arena benchmark, autoregressive language modeling, and ImageNet classification. For instance, Transformer-LS achieves 0.97 test BPC on enwik8 using half the number of parameters than previous method, while being faster and is able to handle 3x as long sequences compared to its full-attention version on the same hardware. On ImageNet, it can obtain the state-of-the-art results (e.g., a moderate size of 55.8M model solely trained on 224x224 ImageNet-1K can obtain Top-1 accuracy 84.1%), while being more scalable on high-resolution images. The source code and models are released at https://github.com/NVIDIA/transformer-ls .