Go To:

Paper Title Paper Authors Table Of Contents Abstract References
Home
Report a problem with this paper

Asynchronous Temporal Fields for Action Recognition

Authors

  • Gunnar A. Sigurdsson
  • S. Divvala
  • Ali Farhadi
  • Abhinav Gupta
  • 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
  • 2017
  • View in Semantic Scholar

Abstract

Actions are more than just movements and trajectories: we cook to eat and we hold a cup to drink from it. A thorough understanding of videos requires going beyond appearance modeling and necessitates reasoning about the sequence of activities, as well as the higher-level constructs such as intentions. But how do we model and reason about these? We propose a fully-connected temporal CRF model for reasoning over various aspects of activities that includes objects, actions, and intentions, where the potentials are predicted by a deep network. End-to-end training of such structured models is a challenging endeavor: For inference and learning we need to construct mini-batches consisting of whole videos, leading to mini-batches with only a few videos. This causes high-correlation between data points leading to breakdown of the backprop algorithm. To address this challenge, we present an asynchronous variational inference method that allows efficient end-to-end training. Our method achieves a classification mAP of 22.4% on the Charades [42] benchmark, outperforming the state-of-the-art (17.2% mAP), and offers equal gains on the task of temporal localization.

1. Introduction

Consider the video shown in Figure 1 : A man walks through a doorway, stands at a table, holds a cup, pours something into it, drinks it, puts the cup on the table, and finally walks away. Despite depicting a simple activity, the video involves a rich interplay of a sequence of actions with underlying goals and intentions. For example, the man stands at the table 'to take a cup', he holds the cup 'to drink from it', etc. Thorough understanding of videos requires us to model such interplay between activities as well as to reason over extensive time scales and multiple aspects of actions (objects, scenes, etc) .

Figure 1. Understanding human activities in videos requires jointly reasoning about multiple aspects of activities, such as ‘what is happening’, ‘how’, and ‘why’. In this paper, we present an end-toend deep structured model over time trained in a stochastic fashion. The model captures rich semantic aspects of activities, including Intent (why), Category (what), Object (how). The figure shows video frames and annotations used in training from the Charades [43] dataset.

Most contemporary deep learning based methods have treated the problem of video understanding as that of only appearance and motion (trajectory) modeling [44, 54, 7 , * Work was done while Gunnar was an intern at AI2. 28]. While this has fostered interesting progress in this domain, these methods still struggle to outperform models based on hand-crafted features, such as Dense Trajectories [57] . Why such a disconnect? We argue that video understanding requires going beyond appearance modeling, and necessitates reasoning about the activity sequence as well as higher-level constructs such as intentions. The recent emergence of large-scale datasets containing rich sequences of realistic activities [43, 64, 61] comes at a perfect time facilitating us to explore such complex reasoning. But what is the right way to model and reason about temporal relations and goal-driven behaviour? Over the last couple of decades, graphical models such as Conditional Random Fields (CRFs) have been the prime vehicles for structured reasoning. Therefore, one possible alternative is to use ConvNet-based approaches [20] to provide features for a CRF training algorithm. Alternatively, it has been shown that integrating CRFs with ConvNet architectures and training them in an end-to-end manner provides substantial improvements in tasks such as segmentation and situation recognition [67, 1, 63] .

Inspired by these advances, we present a deep-structured model that can reason temporally about multiple aspects of activities. For each frame, our model infers the activity cate-gory, object, action, progress, and scene using a CRF, where the potentials are predicted by a jointly end-to-end trained ConvNet over all predictions in all frames. This CRF has a latent node for the intent of the actor in the video and pairwise relationships between all individual frame predictions.

While our model is intuitive, training it in an end-to-end manner is a non-trivial task. Particularly, end-to-end learning requires computing likelihoods for individual frames and doing joint inference about all connected frames with a CRF training algorithm. This is in stark contrast with the standard stochastic gradient descent (SGD) training algorithm (backprop) for deep networks, where we require mini-batches with a large number of independent and uncorrelated samples, not just a few whole videos. In order to handle this effectively: (1) we relax the Markov assumption and choose a fully-connected temporal model, such that each frame's prediction is influenced by all other frames, and (2) we propose an asynchronous method for training fully-connected structured models for videos. Specifically, this structure allows for an implementation where the influence (messages) from other frames are approximated by emphasizing influence from frames computed in recent iterations. They are more accurate, and show advantage over being limited to only neighboring frames. In addition to being more suitable for stochastic training, fullyconnected models have shown increased performance on various tasks [19, 67] .

In summary, our key contributions are: (a) a deep CRF based model for structured understanding and comprehensive reasoning of videos in terms of multiple aspects, such as action sequences, objects, and even intentions; (b) an asynchronous training framework for expressive temporal CRFs that is suitable for end-to-end training of deep networks; and, (c) substantial improvements over state-of-theart, increasing performance from 17.2% mAP to 22.4% mAP on the challenging Charades [43] benchmark.

2. Related Work

Understanding activities and actions has an extensive history [33, 60, 23, 17, 24, 2, 27, 57, 30, 22] . Interestingly, analyzing actions by their appearance has gone through multiple iterations. Early success was with handcrafted representations such as Space Time Interest Points (STIP) [23] , 3D Histogram of Gradient (HOG3D) [17] , Histogram of Optical Flow (HOF) [24] , and Motion Boundary Histogram [2] . These methods capture and analyze local properties of the visual-temporal datastream. In the past years, the most prominent hand-crafted representations have been from the family of trajectory based approaches [27, 57, 30, 22] , where the Improved Dense Trajectories (IDT) [57] representation is in fact on par with state-of-the-art on multiple recent datasets [8, 43] .

Recently there has been a push towards mid-level rep-resentations of video [38, 47, 13, 21] , that capture beyond local properties. However, these approaches still used handcrafted features. With the advent of deep learning, learning representations from data has been extensively studied [14, 15, 45, 58, 53, 54, 25, 7, 62, 56, 41, 3] . Of these, one of the most popular frameworks has been the approach of Simonyan et al. [45] , who introduced the idea of training separate color and optical flow networks to capture local properties of the video. Many of those approaches were designed for short clips of individual activities and hence do not generalize well to realistic sequences of activities. Capturing the whole information of the video in terms of temporal evolution of the video stream has been the focus of some recent approaches [52, 6, 12, 36, 50, 31] . Moving towards more expressive deep networks such as LSTM has become a popular method for encoding such temporal information [49, 4, 66, 51, 59, 42, 65] . Interestingly, while those models move towards more complete understanding of the full video stream, they have yet to significantly outperform local methods [45] on standard benchmarks.

A different direction in understanding comes from reasoning about the complete video stream in a complementary direction -Structure. Understanding activities in a human-centric fashion encodes our particular experiences with the visual world. Understanding activities with emphasis on objects has been a particularly fruitful direction [26, 37, 9, 35, 55] . In a similar vein, some works have also tried modeling activities as transformations [59] or state changes [5] . Recently, there has been significant progress in modelling the complete human-centric aspect, where image recognition is phrased in terms of objects and their roles [63, 10] . Moving beyond appearance and reasoning about the state of agents in the images requires understanding human intentions [16, 32] . This ability to understand people in terms of beliefs and intents has been traditionally studied in psychology as the Theory of mind [34] . How to exactly model structure of the visual and temporal world has been the pursuit of numerous fields. Of particular interest is work that combines the representative power of deep networks with structured modelling. Training such models is often cumbersome due to the differences in jointly training deep networks (stochastic sampling) and sequential models (consecutive samples) [29, 67] . In this work, we focus on fully-connected random fields, that have been popular in image segmentation [19] , where image filtering was used for efficient message passing, and later extended to use CNN potentials [40] .

3. Proposed Method

Given a video with multiple activities, our goal is to understand the video in terms of activities. Understanding activities requires reasoning about objects being interacted Figure 2 . An overview of our structured model. The semantic part captures object, action, etc. at each frame, and temporal aspects captures those over time. On the left side, we show how for each timepoint in the video, a Two-Stream Network predicts the potentials. Our model jointly reasons about multiple aspects of activities in all video frames. The Intent captures groups of activities of the person throughout the whole sequence of activities, and fine-grained temporal reasoning is through fully-connected temporal connections.

Figure 2. An overview of our structured model. The semantic part captures object, action, etc. at each frame, and temporal aspects captures those over time. On the left side, we show how for each timepoint in the video, a Two-Stream Network predicts the potentials. Our model jointly reasons about multiple aspects of activities in all video frames. The Intent captures groups of activities of the person throughout the whole sequence of activities, and fine-grained temporal reasoning is through fully-connected temporal connections.

with, the place where the interaction is happening, what happened before and what happens after this current action and even the intent of the actor in the video. We incorporate all these by formulating a deep Conditional Random Field (CRF) over different aspects of the activity over time. That is, a video can be interpreted as a graphical model, where the components of the activity in each frame are nodes in the graph, and the model potentials are the edges in the graph.

In particular, we create a CRF which predicts activity, object, etc., for every frame in the video. For reasoning about time, we create a fully-connected temporal CRF, referred as Asynchronous Temporal Field in the text. That is, unlike a linear-chain CRF for temporal modelling (the discriminative counterpart to Hidden Markov Models), each node depends on the state of every other node in the graph. We incorporate intention as another latent variable which is connected to all the action nodes. This is an unobserved variable that influences the sequence of activities. This variable is the common underlying factor that guides and better explains the sequence of actions an agent takes. Analysis of what structure this latent variable learns is presented in the experiments. Our model has three advantages: (1) it addresses the problem of long-term interactions; (2) it incorporates reasoning about multiple parts of the activity, such as objects and intent; and (3) more interestingly, as we will see, it allows for efficient end-to-end training in an asynchronous stochastic fashion.

3.1. Architecture

In this work we encode multiple components of an activity. Each video with T frames is represented as {X 1 , . . . , X T , I} where X t is a set of frame-level random variables for time step t and I is an unobserved random variable that represent global intent in the entire video. We can further write X t = {C t , O t , A t , P t , S t }, where C is the activity category (e.g., 'drinking from cup'), O corresponds to the object (e.g., 'cup'), A represents the action (e.g., 'drink'), P represents the progress of the activity {start, middle, end}, and S represents the scene (e.g. 'Dining Room'). For clarity in the following derivation we will refer to all the associated variables of X t as a single random variable X t . A more detailed description of the CRF is presented in the appendix.

Mathematically we consider a random field {X, I} over all the random variables in our model ({X 1 , . . . , X T , I}). Given an input video V ={V 1 , . . . , V T }, where V t is a video frame, our goal is to estimate the maximum a posteriori labeling of the random field by marginalizing over the intent I. This can be written as:

EQUATION (1): Not extracted; please refer to original document.

For clarity in notation, we will drop the conditioning on V and write P (X, I). We can define P (X, I) using Gibbs distribution as:

P (X, I)= 1 Z(V) exp (−E(x, I)) where E(x, I

) is the Gibbs energy over x. In our CRF, we model all unary and pairwise cliques between all frames {X 1 , . . . , X T } and the intent I. The Gibbs energy is:

E(x, I) = i φ X (x i ) Semantic + i φ XI (x i , I) + i,j i =j φ XX (x i , x j ) Temporal , (2)

where φ XX (x i , x j ) is the potential between frame i and frame j, and φ XI (x i , I) is the potential between frame i and the intent. For notational clarity φ X (x i ) incorporates all unary and pairwise potentials for C t , O t , A t , P t , S t . The model is best understood in terms of two aspects: Semantic Figure 3 . Illustration of the learning algorithm, and the message passing structure. Each timepoint that has been processed has a message (Blue highlights messages that have recently been computed). The loss receives a combination of those messages, uses those to construct new messages, and updates the network.

Figure 3. Illustration of the learning algorithm, and the message passing structure. Each timepoint that has been processed has a message (Blue highlights messages that have recently been computed). The loss receives a combination of those messages, uses those to construct new messages, and updates the network.

aspect, which incorporates the local variables in each frame (C t , O t , A t , P t , S t ); and Temporal aspect, which incorporates interactions among frames and the intent I. This is visualized in Figure 2 . We will now explain the semantic, and temporal potentials. Semantic aspect The frame potential φ X (x i ) incorporates the interplay between activity category, object, action, progress and scene, and could be written explicitly as φ X (C t , O t , A t , P t , S t ).

In practice this potential is composed of unary, pairwise, and tertiary potentials directly predicted by a CNN. We found predicting only the following terms to be sufficient without introducing too many additional parameters:

φ X (C t , O t , A t , P t , S t )=φ(O t , P t )+φ(A t , P t )+φ(O t , S t )+ φ(C t , O t , A t , P t )

where we only model the assignments seen in the training set, and assume others are not possible. Temporal aspect The temporal aspect of the model is both in terms of the frame-intent potentials φ XI (x i , I) and frame-frame potentials φ XX (x i , x j ). The frame-intent potentials are predicted with a CNN from video frames (pixels and motion). The pairwise potentials φ XX (x i , x j ) for two time points i and j in our model have the form:

EQUATION (3): Not extracted; please refer to original document.

where µ models the asymmetric affinity between frames, w are kernel weights, and each k (m) is a Gaussian kernel that depends on the videoframes v i and v j . In this work we use a single kernel that prioritises short-term interactions:

EQUATION (4): Not extracted; please refer to original document.

The parameters of the general asymmetric compatibility function µ(x i , x j ) are learned from the data, and σ is a hyper-parameter chosen by cross-validation.

3.2. Inference

While it is possible to enumerate all variable configurations in a single frame, doing so for multiple frames and their interactions is intractable. Our algorithm uses a structured variational approximation to approximate the full probability distribution. In particular, we use a mean-field approximation to make inference and learning tractable. With this approximation, we can do inference by keeping track of message between frames, and asynchronously train one frame at a time (in a mini-batch fashion).

More formally, instead of computing the exact distribution P (X, I) presented above, the structured variational approximation finds the distribution Q(X, I) among a given family of distributions that best fits the exact distribution in terms of KL-divergence. By choosing a family of tractable distributions, it is possible to make inference involving the ideal distribution tractable. Here we use Q(X, I) = Q I (I) i Q i (x i ), the structured mean-field approximation. Minimizing the KL-divergence between those two distributions yields the following iterative update equation:

Q i (x i ) ∝ exp φ X (x i ) + E U ∼Q I [φ XI (x i , U )] + j>i E U j ∼Q j [φ XX (x i , U j )] + j

where Q i is marginal distribution with respect to each of the frames, and Q I is the marginal with respect to the intent. An algorithmic implementation of this equation is as presented in Algorithm 1.

Algorithm 1 Inference For Asynchronous Temporal Fields

1: Initialize Q Uniform distribution 2: while not converged do 3:

Visit frame i 4:

Get j>i E U j ∼Q j [φ XX (x i , U j )] 5: Get j

while not converged do 8:

Update Q i and Q I using Eq. 6 9:

Send

E U ∼Q i [φ XX (x, U )] 10: Send E U ∼Q i [φ XX (U, x)] 11: Send E U ∼Q i [φ XI (U, I)]

Here 'Get' and 'Send' refer to the message server, and f (x) is a message used later by frames in the same video. Figure 4 . Evolution of prediction with increasing messages passes. The first row shows the initial prediction for the category tidying with a broom without any message passing, where darker colors correspond to higher likelihood, blue is then an increase in likelihood, and brown decrease. In the first message pass, the confidence of high predictions gets spread around, and eventually increases the confidence of the whole prediction.

Figure 4. Evolution of prediction with increasing messages passes. The first row shows the initial prediction for the category tidying with a broom without any message passing, where darker colors correspond to higher likelihood, blue is then an increase in likelihood, and brown decrease. In the first message pass, the confidence of high predictions gets spread around, and eventually increases the confidence of the whole prediction.

distributes them accordingly when requested. In practice, this could be implemented in a multi-machine setup.

3.3. Learning

Training a deep CRF model requires calculating derivatives of the objective in terms of each of the potentials in the model, which in turn requires inference of P (X, I|V ). The network is trained to maximize the log-likelihood of the data l(X) = log I P (x, I|V ). The goal is to update the parameters of the model, for which we need gradients with respect to the parameters. Similar to SGD, we find the gradient with respect to one part of the parameters at a time, specifically with respect to one potential in one frame. That is, φ i X (x) instead of φ X (x). The partial derivatives of this loss with respect to each of the potentials are as follows:

∂l(X) ∂φ i X (x) = 1 x=x − Q i (x) (7) ∂l(X) ∂φ i XI (x,Î) = exp j φ XI (x j ,Î) I exp j φ XI (x j , I) 1 x=x − Q i (x)Q I (Î) (8) ∂l(X) ∂µ i (a, b) = j>i 1x=ak(v i , v j ) − Q i (x) j>i Q I (b)k(v i , v j ) + j

where φ i X (x) and φ i XI (x,Î) is the frame and frame-intent potentials of frame i, and we usex to distinguish between the labels and variables the derivative is taken with respect to. µ i (a, b) are the parameters of the asymmetric affinity kernel with respect to frame i, and 1 x=x is a indicator variable that has the value one if the ground truth label corresponds to the variable. Complete derivation is presented in the appendix. These gradients are used to update the underlying CNN model. These update equations lead to the learning procedure presented in Algorithm 2. for each example in mini-batch do 4:

Sample frame v ∈ V ⊆ V 5:

Get incoming messages 6:

Update Q i and Q I 7:

Find gradients with Eq. 7-9 8:

Backprop gradients through CNN 9:

Send outgoing messages to calculate each partial gradient. This shares ideas with contrastive divergence [11, 39] . Given a single video at test time, we visualize in Figure 4 how the predictions changes as the distribution converges with multiple messages passes.

Message Passing The key thing to note is all the incoming messages are of the form M (z)= j f j (z) where f j is some function from node j; for e.g.,

M (z) = j E Uj ∼Qj [φ XI (U j , z)] = j f j (z)

from Algorithm 1. We use the following approximation during training:

EQUATION (10): Not extracted; please refer to original document.

where

d ∈ [0, 1]

is a discount factor, h is a hyperparameter, and J(•) is an ordering of the messages in that video based on the iteration in which the message was computed. The messages are a weighted combination of stored messages.

4. Experimental Results And Analysis

We analyzed the efficacy of our model on the challenging tasks of video activity classification and temporal localization. In addition, we investigated the different parts of the model, and will demonstrate how they operate together. Dataset Recent years have witnessed an emergence of large-scale datasets containing sequences of common daily activities [43, 64, 61] . For our evaluation, we chose the Charades dataset [43] . This dataset is a challenging benchmark containing 9,848 videos across 157 action classes with 66,500 annotated activities, including nouns (objects), verbs (actions), and scenes. A unique feature of this dataset is the presence of complex co-occurrences of realistic humangenerated activities making it a perfect test-bed for our analysis. We evaluate video classification using the evaluation criteria and code from [43] . Temporal localization is evaluated in terms of per-frame classification using the provided temporal annotations. Implementation details We use a VGG16 network [46] with additional layers to predict the model potentials (Figure 5) . We train both a network on RGB frames, and stacks of optical flow images, following the two-stream architecture [45] . The main challenge in training the network is the increase in the output layer size. For the larger potentials, Figure 5 . The VGG-16 variant predicts the potentials for both RGB and Flow. The network predicts the values of all potentials except one (in this figure we group the frame potentials φX into one layer for clarity). The model is trained end-to-end by passing gradients from the Asynchronous Temporal Field through the network.

Figure 5. The VGG-16 variant predicts the potentials for both RGB and Flow. The network predicts the values of all potentials except one (in this figure we group the frame potentials φX into one layer for clarity). The model is trained end-to-end by passing gradients from the Asynchronous Temporal Field through the network.

we used the following structure to go from fc7 to φ XI : Linear layer (4096 to 100), ReLU, Dropout, Linear layer (100 to the potential values). The input to the RGB network is an image of size 224×224×3 where we crop random location, size, and aspect ratio. We use data augmentation with color jitter and PCA lighting noise. The RGB network was pretrained on ImageNet. The input to the Flow network is a stack of 10 consecutive optical flow frames at 24 FPS starting with the current frame. Since each optical flow has two channels, the input size is 224×224×20 as in [45] . The Flow network was pretrained on UCF101 [48] as in Sigurdsson et al. [43] , and random cropped in the same way as RGB.

We follow the training setup in Charades [43] and consider a frame to have one activity label at a time. Even so, our method is still able to reason about other activities in the video. Convergence of the model is evaluated using the approximate distribution Q i (X) at each frame. The Charades dataset has the property that scenes were chosen at random for each sequence of activities. For this reason, we found reasoning about scenes to reduce the performance, and the weight of that term was lowered in the model.

To obtain annotations for action progress p t , we split each activity annotation into three equally sized parts. All layers of the network are trained with a batch size of 240 and a learning rate of 10 −3 (RGB), 10 −5 (Flow). Learning rate was reduced by a factor of 10 every 30k iterations for RGB, and every 140k iterations for Flow. The value of the message decay parameter d was set to d = 0.9, and the standard deviation σ in (4) was set to 6.25 sec (150 frames).

For testing, we sampled 25 equally spaced frames from the video and synchronously pass messages between the frames until convergence (10 message passes). The predictions of the RGB and Flow networks are combined in a probabilistic fashion by multiplying their probabilistic predictions for each class. More implementation details may be found in the appendix. The networks were implemented in Torch, and the code is available on project page. Diverse batches As highlighted in Section 1, the standard Iterations Figure 6 . Convergence of our method compared to other methods that capture temporal structure. Our asynchronous training method contains more diverse batches, has faster and more stable convergence, and reaches higher accuracy on the test set.

Figure 6. Convergence of our method compared to other methods that capture temporal structure. Our asynchronous training method contains more diverse batches, has faster and more stable convergence, and reaches higher accuracy on the test set.

way of sampling batches for temporal models results in high correlation between data points leading to a breakdown of the SGD. To understand the importance of having many diverse examples from multiple videos, we compare the convergence of our method to two alternatives using homogeneous batches: CNN+LSTM from Ng et al. [66] , and a synchronous version of our method, where each batch contains full videos (only three videos fit into each mini-batch).

We do synchronous message passing until convergence before calculating gradients for backprop. Figure 6 shows that our asynchronous training method, containing more diverse training batches, has faster and more stable convergence.

4.1. Video Classification

Given a video, the task here is to verify whether it contains one or several of the 157 activity categories. Classification accuracy is measured with the standard mean average precision (mAP) criterion, where a prediction is given for each video. This task has been shown to be highly challenging, with the state-of-the-art non-ensemble methods reaching an mAP of only 17.2%, particularly as each video in this dataset has a sequence of multiple fine-grained activities with a real-world long-tailed activity distribution.

We trained our models using the provided training split following the procedure outlined in Section 3. To make predictions for the whole video, we marginalize out everything except the activity category for 25 equidistant frames in the video. The score for each activity category is the maximum across all frames following the setup from [43] . In our analysis, we include the provided non-ensemble baselines from [43] as well as the following additional baselines:

Two-Stream++. We reimplemented the network described in [43] , which follows Simonyan et al. [46] , with the same parameters. We added data augmentation and finetuned all layers of the network. The performance of only the RGB stream is included (RGB++). We also consider Two-Stream Extended which is the same network, but the Flow network was trained for 25 times more iterations than the RGB network (two weeks of computation on a Titan X Figure 7 . The classes with the highest positive and negative difference between our method and Two-Stream (no structure). Our method does better on many classes, without doing much worse on any. In particular, activities that have temporal structure, such as Opening/Closing a refrigerator have significantly higher performance, since our model can reason jointly about those. GPU). Combined with the augmentation, we found this to non-trivially increase the accuracy. Two-Stream+LSTM. We followed the method outlined in [66] to jointly train a LSTM on top of the two-stream network. We trained both an RGB and an Optical Flow network using the same setup from [43] . The trained networks from Two-Stream++ were used to initialize the models. Table 1 displays the accuracy obtained by our method along with the baselines. Our proposed approach obtains an mAP of 22.4% substantially outperforming the Twostream Extended baseline at 18.6% mAP, and the IDT baseline at 17.2%. Our method reasons over significantly larger timescales and multiple aspects of the activities. To ascertain this, we highlight in Figure 7 , the activity classes with the highest positive and negative difference between our method and the Two-Stream network. It is interesting to note that two of those activities are opening and closing a refrigerator, that arguably have a significant causal structure (an open refrigerator was opened at some point), which our model harnesses to significantly increase the accuracy.

Figure 7. The classes with the highest positive and negative difference between our method and Two-Stream (no structure). Our method does better on many classes, without doing much worse on any. In particular, activities that have temporal structure, such as Opening/Closing a refrigerator have significantly higher performance, since our model can reason jointly about those.
Table 1. Video classification results on Charades [43]. The left shows the published baselines from [43] and the right show additional new baselines. Our proposed approach outperforms all competing methods on this dataset.

Ablation studies To study the contribution of different model parts, we also train ablated versions of our model separately choosing the best hyperparameters for each version. In addition to our model with only RGB or Flow, we also consider dropping φ XX (i.e., no sequential informa- Table 2 . Temporal localization results (mAP %) on the Charades [43] dataset. Our proposed method outperforms the LSTM model, and is also more tractable to train at a large-scale. tion), φ XI (i.e., no intent), both (i.e., only semantic information), and further dropping φ X (i.e., dropping all structure). Figure 8 shows that semantic reasoning improves over the baseline. Further, while both φ XI and φ XX capture temporal information, they are complementary.

Figure 8. Ablation analysis for our proposed model. Y-axis is video classification mAP %. Each factor helps in improving the overall model performance. φ(P ) indicates dropping the ‘progress’ term within the semantic factor φX .
Table 2. Temporal localization results (mAP %) on the Charades [43] dataset. Our proposed method outperforms the LSTM model, and is also more tractable to train at a large-scale.

4.2. Temporal Localization

To measure the ability of the methods to temporally localize and understand when exactly activities happen, we adapt the benchmark of [43] to evaluate with the same mAP metric but on individual frames. That is, instead of having a single prediction per video, evaluation is now split into 25 equidistant timepoints having zero or more activities, and the models make a prediction for each of those * . We find this way of evaluating localization robust to annotation ambiguity, and informative for challenging datasets. All hyperparameters were kept equal between localization and classification experiments. All baselines are run on 75 frames across the video, and then every third frame selected for a total of 25 frames. We also considered methods with postprocessing where the model predictions for the 75 frames are averaged across 30 frames to obtain more spatial consistency, and then 25 frames selected as before. Table 2 shows that our method outperforms the alternatives, including the LSTM model which has been shown to be a powerful temporal modeling tool, but challenging to train on top of a two-stream network due to correlations between consecutive samples. These results demonstrate the our method is tractable way of training end-to-end structured models to understand activities. Interestingly, our method still benefits from adding post-processing, significantly more than the LSTM baseline, likely since our method is reasoning on larger time-scales. This suggests Category: Sitting in a chair Category: Reading a book Category: Holding a book Action: sit Action: hold Object: book Figure 9 . Model predictions for a sample video. We see the interplay between categories, objects and actions over time. For example, model becomes confident about the action sit early, which aids the understanding of Sitting in a chair once the chair becomes visible, and helps predicting Reading a book. Darker colors represent higher likelihood, and we average predictions to correspond to each frame.