Only Time Can Tell: Discovering Temporal Data for Temporal Modeling

TL;DR

This paper introduces a new dataset of action classes that require temporal information for recognition, addressing the bias in current datasets and providing a benchmark for evaluating temporal understanding in video models.

Contribution

The authors propose a human-annotated methodology to identify truly temporal classes and create a dataset that isolates temporal information, facilitating better evaluation of temporal modeling capabilities.

Findings

Models perform worse on temporal classes without temporal cues.

Training on the temporal dataset improves generalization to unseen classes.

Current datasets contain many static classes that do not require temporal understanding.

Abstract

Understanding temporal information and how the visual world changes over time is a fundamental ability of intelligent systems. In video understanding, temporal information is at the core of many current challenges, including compression, efficient inference, motion estimation or summarization. However, in current video datasets it has been observed that action classes can often be recognized without any temporal information from a single frame of video. As a result, both benchmarking and training in these datasets may give an unintentional advantage to models with strong image understanding capabilities, as opposed to those with strong temporal understanding. In this paper we address this problem head on by identifying action classes where temporal information is actually necessary to recognize them and call these "temporal classes". Selecting temporal classes using a computational method would bias the process. Instead, we propose a methodology based on a simple and effective human annotation experiment. We remove just the temporal information by shuffling frames in time and measure if the action can still be recognized. Classes that cannot be recognized when frames are not in order are included in the temporal Dataset. We observe that this set is statistically different from other static classes, and that performance in it correlates with a network's ability to capture temporal information. Thus we use it as a benchmark on current popular networks, which reveals a series of interesting facts. We also explore the effect of training on the temporal dataset, and observe that this leads to better generalization in unseen classes, demonstrating the need for more temporal data. We hope that the proposed dataset of temporal categories will help guide future research in temporal modeling for better video understanding.

Figures

Tables

Table 1. Table 1 : Evaluation of popular action recognition networks. The ranking of methods changes when we use the traditional score and when we use the proposed temporal score.

Network	R2D (RGB)	R2D (Flow)	R2+1D (RGB)	R2+1D (Flow)	I3D (RGB)	I3D (Flow)	R3D (RGB)	R3D (Flow)
Temporal Score ($\nabla$ %)	1.13	8.74	4.21	9.50	3.60	6.29	5.19	7.36
Traditional Accuracy (%)	31.83	22.23	44.58	41.92	51.53	43.35	50.48	45.50

Equations

$$\text{Score}(N)={\overline{\text{Acc}}}_N(T)-{\overline{\text{Acc}}}_N(S),$$

Full text

1 Introduction

Temporal information has been shown to be important for recognition. In the biological setting, Johansson’s point-light experiments [15] show that moving dots alone contain enough information to understand actions, when the points-lights are placed at the joints. Similarly, Heider and Simmel [9] find that simple figures (like squares or triangles) in motion can convey rich story information, including intention and emotions like anger, joy. Figure 2(a) shows two frames from these experiments; neither the action nor a story can be recognized easily in the absence of temporal information.

In computational settings, temporal features (optical flow, spatio-temporal convolutions, or both) have been shown to be useful for video understanding. For some datasets and models, using optical flow as input actually yields higher accuracy than using RGB images [25, 2, 29]. However, a single image is often informative enough to predict the label with good confidence. This leads to the question when do we need temporal information to understand an action? Figure 2(b) shows some examples from Kinetics [17], where the action class is easy to guess. Furthermore, in recent two-stream models such as I3D [2] fusing the predictions of the stream operating on optical flow (the temporal stream) to that operating on RGB frames (the spatial stream) produces rather small gains. Do these observations mean that given the image content, motion information is redundant? We argue that actually these observations do not reflect the importance of motion in video understanding, but rather the distribution of categories in some of the video datasets. Based on this hypothesis, we propose to examine some of the most widely used datasets, and identify classes where temporal information is necessary to recognize them. We call these “temporal classes." Our analysis is based on human perception rather than on the performance of computational models as we do not want to tie the definition of temporal categories to a specific model’s performance. In particular, given a collection of videos , we compare human performance on a video with shuffled frames and on the same video with the original ordering of frames. We use this difference as a cue for how necessary temporal information is to recognize a class. We then select the classes where this difference is largest. We use these categories as a benchmark to measure how well different architectures perform. Our analyses provides interesting observations such as that the ranking of models on the temporal classes is different compared to that on the full datasets and that temporal models trained on temporal classes have stronger generalization performance. In summary, we make the following contributions:

• •

A methodology to discover the importance of temporal information in action classes based on human perception. This allows us to untie the definition of temporal relevance from the performance of existing computational methods.

• •

The Temporal Dataset, resulting from using our proposed methodology on current video classification datasets. We identify the set of categories containing more temporal information and what they have in common, and examine their computational properties. We will release all of the collected human judgments for further analysis by the community.

• •

We benchmark current video models, and observe that the ranking changes with the inclusion or omission of temporal categories, revealing hidden biases and strengths. This also suggests that the distribution of the videos in the temporal categories actually has different structure.

• •

We use the temporal categories for training temporal models. We observe that training on the Temporal Dataset produces stronger temporal features. We also observe that training on this dataset leads to a better ability to generalize in the transfer learning setting of training on “seen" categories and testing on “unseen" categories.

We hope that our proposed “meta-dataset" of temporal categories will help guide future research in temporal modeling for better video understanding.

2 Related Work

Temporal Modeling. Over the last years video understanding has made rapid progress. While modeling spatial information on videos has largely been addressed leveraging advancements in image understanding methods, modeling temporal information has sprouted new and creative techniques. These techniques can be broadly categorized in three families: two-stream, temporal convolutions, and recurrent methods. The first, pioneered by Simonyan and Zisserman [25], consists on having two parallel networks: the spatial stream aimed at modeling image content, which takes RGB as input, and a temporal stream aimed at modeling motion content, which takes a small number of neighboring flow fields as input. The predictions of the two streams are combined at the end. Different variants of this general framework arise using different architectures for the streams [30]. These techniques are described as frame-based, since they do not explicitly model inter-frame information. Temporal convolutions, like C3D [28] extend the success of 2D convolutions in image understanding to include 3D convolutions across frames that are adjacent in time. Finally, recurrent convolutional networks [27, 31, 5] have been used for learning video representations, most often by applying a 2D CNN as a feature extractor on the individual frames and using a recurrent network on the CNN features to model temporal information. Some of the latest methods actually combine the two-stream approach with the temporal convolution approach like the I3D architecture [2] or the R2+1D [29]. In this paper we will focus on this hybrid family of methods, that include two streams and temporal convolutions, since they will allow us to experiment and compare the impact of testing and training on temporal data when using flow vs using images, and using frame-based models vs using temporal convolutions.

Video Datasets. Progress in video understanding has been greatly propelled by the growth of video datasets in data scale and richness of taxonomy. Two of the first datasets in this area were UCF-101 [26], HMDB-51 [19], which have been widely studied since the early 2010’s. These are single-label videos of human actions in 101 classes and 51 classes, respectively. The first large-scale dataset to appear was Sports-1M [16], which provides 1M videos of sport actions in 487 classes. Although the labels are predicted by tagging algorithms and thus noisy, Sports-1M has consistently improved the quality of models pre-trained on it. In the last few years, researchers have pushed video classification to more challenging scenarios by using human annotation to collect large-scale datasets with different characteristics. For example, the Something-Something [10] dataset contains categories of generic actions independent of objects, with the goal of obtaining models that reason over physical aspects of actions and scenes as opposed to high-level action concepts. Kinetics [17] includes 400 categories of human actions in different environments, combining the success case of UCF101 with large-scale. Unlike past datasets that focus on third person videos, Charades [24] focuses on egocentric first-person videos that happen around the house. These datasets are often created by first deciding on a taxonomy of classes and then looking for videos containing instances of such classes, or recording them [21]. This process may bias the distribution of videos based on the search engine, and thus SOA [22] proposes a dataset where the videos are chosen first, and then labeled using the taxonomy of classes that naturally arises. While this is a more expensive labeling process, it yields videos often with multiple labels, and datasets with less biased label taxonomy. Other recent datasets also extend the types of video tasks to action localization(e.g. THUMOS 14 [14], ActivityNet [1], AVA [11]), video object detection (e.g. VLOG [8]), and video prediction (e.g. Epic-kitchens [4]). Despite the tremendous efforts of dataset collection, it is not well studied yet what action classes substantially require temporal modeling to recognize as opposed to still appearance modeling.

Temporal Ordering and Frame Shuffling. Temporal ordering in the video is arguably important for video understanding. A rich set of methods have explicitly exploited it for video recognition, including ranking pooling [7, 3], and RNN-based methods [6, 27]. Furthermore, Misra et al. [20] use it as a self-supervised task to learn better temporal models, where a network is trained to predict the original ordering of a shuffled set of frames. Other recent work has explored the effect of shuffling to measure the robustness of a network to temporal disarray [23, 13], observing that even networks using temporal convolutions can be surprisingly insensitive to shuffling frames within clips. Finally, perceptual studies also have investigated the effect of shuffling in space and time [18]. In contrast, in this work, we conduct a human perception study, and use frame shuffling to identify action classes where the action recognition performance of human is substantially compromised by the randomly shuffled temporal ordering.

3 The Temporal Dataset

In this section we describe the process to discover the categories for which temporal information is important, which will be part of what we call the Temporal Dataset. We first describe the perceptual test in detail and then discuss our findings.

3.1 Perceptual Test to Discover Temporal Action Classes

Motivation.

How can we measure if temporal information is necessary to identify an action in a video? We could, for example, compare the accuracy of a method using the entire video as input to the accuracy using a single frame. However, going from the full video to a single frame eliminates not only the temporal information, but also some appearance information, since we would be excluding most of the frames. Further, recent studies have shown [13] that some frames make recognition easier than others, raising a more complicated “frame selection problem". Instead we maintain all image information and removes all temporal information by shuffling frames in time, and using the resulting video as input. We could then choose an action recognition network and observe the performance drop with and without temporal shuffling. However, this approach would tie the selection of temporal classes to a specific computational model. This is risky for two reasons. First, it conflates the inherent complexity of identifying a class and the temporal information. Second, we do not know the architecture’s ability to represent temporal information in the first place. Our intent here is not to find classes where a specific model does poorly, but to define a “meta-dataset" of temporal classes that will spur innovation in temporal modeling. Based on these arguments we propose to show the shuffled videos to human subjects and to define temporal relevance of a class based on the drop in accuracy in human recognition performance when the video frames are shuffled in time.

Stimuli.

We empirically found that watching videos where frames are shuffled in time can be discomforting. It is hard for humans to attend to so much change in the scene. Even if we slow down the frame rate there is a fundamental problem: we are simply replacing the true temporal information with random temporal information; the human perceptual system will still attempt to connect the frames in time. For this reason we render the video on a canvas containing space for two frames side by side. At any time we black out one half (either the left or the right half) of the canvas and render the current frame in the other half. The next frame is then shown in the half that was previously blacked out. This left-right alternation eliminates any perception of motion (real or random), producing a slide-show effect where image information is retained but temporal information is removed. Video examples are included in the supplementary material.

Task.

Action recognition networks are typically trained for multi-class classification, where given an input video, they classify the action in it. Ideally we would like to expose human annotators to the exact same task. However, this is difficult in practice, since the number of classes is in the order of hundreds. Annotators would have a hard time remembering all the classes in the taxonomy. Another option is to ask the binary question “is X action happening in this video?". The problem there is that we only have positive examples, and generating compelling negative examples is not trivial. Instead, for each video we show as possible answers only classes within a semantic group. For example, if a video has the label “barbecueing", then we will show as possible answers other cooking categories, which includes classes like “making a sandwich", since they are more likely to be visually similar. 111These semantic groups are provided in the Kinetics dataset, but for the Something-something dataset we design them manually, using the verb as guidance. This restricts the options that we give to the annotator, from the order of hundreds to 5-15 options. While this may be considered making the task easier, in practice it would be rare that categories across groups would be accidentally mistaken for each other.

Datasets.

We discover temporal classes from the most relevant action classification datasets. We use Kinetics [17] (400 classes, 280K videos), for being the most widely used set. We also use Something-Something [10] (174 classes, 100K videos) because it was specifically designed to be action-centered and independent of the objects appearing in the scene which makes it particularly relevant to our study.

Selection of Temporal Classes.

We show between 15 and 30 videos per class to different human annotators. For each class $i$ we compute the average accuracy of the annotators when they see the videos shuffled $\textrm{Acc}_{s}(i)$ and for the control $\textrm{Acc}_{c}(i)$, when the videos are not shuffled. We select the groups where the control accuracy is larger than $70\%$ (e.g., the classes names are sufficiently clear) and the drop in accuracy $\textrm{Acc}_{c}(i)-\textrm{Acc}_{s}(i)$ is larger than a threshold ($40\%$ in Kinetics, and $60\%$ in Something-something, since performance drops have different statistics). We also include those action classes that have been the source of confusion more than $20\%$ of the times, since they are necessary to actually observe the confusion (e.g. plugging and unplugging, opening and closing, etc).

3.2 What Actions Actually Require Temporal Information?

In total, we discover 50 temporal classes, where 32 come from the Kinetics datasets and 18 from Something-something. We call this subset the Temporal Dataset. The total amount of videos is 35045, where 32081 are for training and 2964 for testing.

In Table 1 we show the actions where accuracy drops most when we remove temporal information, by group and by class, respectively. As expected, temporal information is more important when the scene or objects in are not discriminative for the action recognition task, for example in videos showing different forms of dancing, or different skiing styles. These are classes where the background, clothing, and objects tend to be fairly constant across categories or uninformative for action recognition. We also observe that facial actions are particularly sensitive to frame shuffling, making actions like yawning and sneezing difficult to tell apart without temporal information.

Figure 3 shows the histogram of performance drop of all classes after temporal shuffling. The proportion of classes that require temporal analysis is small for both datasets. Something-Something shows a larger proportion of categories requiring temporal information for discrimination. This is not surprising since this dataset was designed precisely to avoid strong correlations between object categories and action classes. Perhaps what may be more surprising is that still many of the classes in such dataset can be correctly identified from shuffled frames. This highlights the need for the proposed methodology: it is difficult to design a dataset of actions that require strong temporal modeling for accurate recognition, while it is much easier to discover temporal classes using our simple perceptual test.

4 Analysis of the Temporal Dataset

4.1 Experimental Details

Datasets.

We use two sets of classes in our experiments: Temporal-50, which is the set of 50 temporal action classes where performance decreases most as in Sec. 3. And Static-50, the 50 classes where human accuracy decreases least, also as measured in Sec. 3. The number of videos in this set is comparable to that of Temporal-50. While there is a larger portion of classes where accuracy decreases a negligible amount, we choose 50 at random to keep the number of classes and videos comparable to Temporal-50. Note that both sets contain the same proportion of classes from Kinetics, and Something-something, to factor out the dataset bias component.

Architectures.

We choose a suite of architectures that are top-performing and most widely used. They include some that use temporal information (e.g. through optical flow or temporal convolution) and others that do not. In particular, we use: R2D which is a frame-based model, that uses the ResNet [12] architecture with 18 layers, and it is trained with a cross-entropy loss; R3D [28] which is similar to R2D, but instead of processing each frame separately it includes convolutions in the temporal dimension; I3D [2] which is also a 3D architecture, that extends inception to use convolutions in the temporal dimension; and R(2+1)D which is a recent architecture [29] that achieves state-of-the-art accuracy. The main difference compared to R3D is that R(2+1)D factors the 3D convolutions into 2D filters in space and 1D filters in time. This separability helps the optimization, making this the strongest of the models we use.

Training and Evaluation Details.

Unless otherwise specified, all training experiments are performed with the same parameters: 75 epochs, from which 10 are warm-ups and the remaining 65 follow a cosine function, where the base learning rate is 0.0025 and gamma is 0.0001, on 4 GPUs at a time. The batch size is 16. We use multiple crops of size 112$\times$112. All models are trained from scratch. Unless otherwise specified, the reported evaluation of models is done using a clip size of 16 frames. We report only the video-level top-1 accuracy in the main paper for brevity.

4.2 Computational Properties of the Temporal Dataset

In this section we perform a series of computational sanity checks, that help validate the nature of the Temporal and Static datasets. This is, that the Temporal Dataset actually captures the classes where temporal information matters, and the Static Dataset captures those where temporal information is redundant. We argue that a randomly chosen set of classes is very unlikely to satisfy all our sanity checks.

Temporal and Static sets are statistically different.

The first sanity check is to make sure that the behavior of networks is different in the Temporal and Static Datasets in a statistically significant manner. Given a network $N$, we compute the per-class accuracy of the network $\textrm{Acc}_{N}(i)$ of all classes $i$. We now compute all accuracies of the classes in the Temporal Dataset $T$, as $A=\{\textrm{Acc}_{N}(i)\;|\;i\in T\}$. We also compute the accuracies in the Static Dataset $S$, which is the set $B=\{\textrm{Acc}_{N}(i)\;|\;i\in S\}$. We use the Kolmogorov-Smirnov test to measure if $A$ and $B$ are sampled from the same distribution, for multiple networks $N$ (R2+1D and R2D). We observe that the null hypothesis can be rejected with $p<0.1$, suggesting that the behavior of the networks on the datasets $T$ and $S$ are sampled from different distributions. As a control experiment, we randomly sample two sets $T^{\prime}$ and $S^{\prime}$, and observe that $p>0.4$. We conclude that the behavior of networks is statistically different in the Temporal and Static datasets.

A network’s ability to capture temporal information is correlated with its accuracy in the Temporal Dataset.

Recall that one of our goals with the creation of the Temporal Dataset is to use it as a benchmark to measure a network’s ability to capture temporal information. In this sanity check we use pairs of networks where we do know their relative abilities to capture temporal information. In particular we use the pair R2D and R2+1D, which are identical except for the fact that R2D is frame-based and R2+1D includes temporal convolutions. Thus it is safe to say that R2+1D has a stronger ability to represent temporal information than R2D. We also use the pair RGB and flow. If the Temporal Dataset actually contains videos where temporal information is necessary, we should observe that R2+1D has an advantage over R2D on the Temporal Dataset, and so does using flow as input over using RGB. Of course, R2+1D has larger capacity, and it will typically always outperform R2D. We use the accuracy on the Static Dataset to control for the capacity of the networks, and measure the difference between the accuracy in the Temporal Dataset and the Static Dataset.

Results are shown in Fig. 5. The accuracy of the R2D network using RGB as input is similar in the Temporal and the Static Datasets (Fig. 5(a)). However, when we use the R2+1D architecture, which adds temporal convolutions, the network does better on the Temporal than the Static set. This is consistent with our hypothesis, that networks that model temporal information better perform better in the Temporal Dataset. Similarly, we observe better performance on the Temporal Dataset when using flow as input, than when using RGB as input. This difference is more clearly observed in Fig. 5(b), which shows the correlation between the networks’ ability to capture temporal information, and its gain in the Temporal Dataset, as expected. At the same time, we do not observe this pattern in the control experiment of sampling subsets of classes at random. With this we conclude that the relative gain in the Temporal Dataset compared to the Static Dataset is a good measure for evaluating a network’s temporal capabilities.

5 Temporal Dataset as a Benchmark

In this section we propose a score to evaluate a network’s ability to capture temporal information based on the experiments from Sec. 4.2. We use this score to rank some of the most recent and widely used networks. In particular, we propose computing the relative gain of the network on the temporal classes, compared to static classes:

[TABLE]

where $\textrm{Acc}_{N}(T)=\{\textrm{Acc}_{N}(i)\mid i\in T\}$ and $\textrm{Acc}_{N}(S)=\{\textrm{Acc}_{N}(i)\mid i\in S\}$. Results of evaluating a suite of popular methods are shown in Table 1.

We make several observations. First, using optical flow shows to be much more useful. This reveals that the perception that optical flow is not useful in recent datasets is a more a product of the task (i.e., solving Kinetics) than the true nature of the visual world. Second, we observe that the top-performing method (I3D) according to the traditional accuracy, scores rather poorly in the temporal score. This is quite possibly a direct consequence of the training process using inflation based on image filters learned from ImageNet. While this process is useful for action recognition, it biases the network to excel in static classes. Since R2+1D and R3D are not pre-trained on Imagenet, their overall performance on the datasets is lower, but they do shine on the temporal classes.

6 Effect of Training on the Temporal Dataset

We also explore the effect of training using the Temporal Dataset. In particular, we are interested in observing the behavior of the training in the absence of discriminative image information. In principle, this should force the network to learn stronger temporal features, which may help generalization. For the experiments we use the R2+1D arquitecure with RGB as input.

Training on the Temporal Dataset avoids image bias and generalizes better.

In this experiment we measure the ability of a network to generalize to unseen classes when the network has been trained on temporal classes, and on static classes. We take each of the two datasets, which contain 50 classes, and we train on 40 of them, and leave the other 10 as unseen classes to test. After training a network on 40 classes, we freeze all weights except for the last layer, which we change to map to 10 classes, and we finetune it. Results are shown in Fig. 6 (a). We would expect that training on temporal classes would be better for testing on temporal classes, and that training on static would be better for testing on static. However, we find that testing on static shows on par performance for both training processes. We hypothesize that in the absence of discriminative image cues, the network cannot “cheat" and is forced to learn better temporal information, which is useful both for static and temporal classes. More importantly, we find that training on temporal performs much better testing on temporal, which shows the need of temporal data to learn good temporal features.

Training on the Temporal Dataset produces features more sensitive to temporal changes.

Shuffling frames at test time has been noted to not harm performance dramatically [23]. This observation reveals that features are not very sensitive to changes in temporal structure. Does training on the Temporal Dataset produce more temporally-aware features? We measure the effect of shuffling frames on networks trained on static and temporal classes. Results are shown in Fig. 6 (b). We observe that the performance of a network trained on temporal classes drops much more than one trained on static classes. This is consistent both when shuffling at the video level and the clip level. We argue that training on temporal classes leads to larger sensitivity to temporal ordering, and therefore stronger temporal features.

7 Conclusion

We have presented a methodology to discover the relevance of temporal information in action classes based on human perception. We use it to identify categories that contain more temporal information and we called them temporal classes. We use this set of classes to benchmark current video models, revealing a network’s ability to model temporal information. We also used these classes to for training and observe that boosts accuracy on the temporal categories, even on unseen classes. We hope that the proposed “meta-dataset" of temporal categories will help guide future research in temporal modeling for better video understanding.222Answers: “sneezing” (left), and “crying” (right).

Appendix

Temporal classes indices:

From Kinetics: [30, 34, 41, 45, 63, 75, 105, 147, 148, 152, 173, 177, 206, 222, 228, 230, 235, 277, 295, 301, 302, 307, 308, 309, 310, 322, 325, 349, 357, 358, 376, 395]

From Something-something: [1, 6, 10, 12, 26, 34, 39, 45, 47, 56, 60, 63, 67, 70, 77, 81, 95, 150]

Static classes indices:

From Kinetics: [18, 20, 23, 26, 53, 57, 80, 88, 91, 95, 99, 113, 126, 139, 181, 182, 183, 186, 200, 201, 210, 218, 226, 227, 255, 268, 298, 354, 366, 388, 397, 398]

From Something-something: [9, 15, 42, 50, 58, 61, 63, 71, 78, 90, 93, 106, 109, 114, 121, 130, 138, 165]

Random sets indices:

From Kinetics: [7, 8, 27, 30, 41, 44, 65, 73, 80, 86, 112, 113, 118, 125, 137, 146, 158, 189, 203, 204, 205, 206, 208, 210, 228, 231, 254, 355, 362, 377, 385, 399]

From Something-something: [0, 2, 5, 18, 27, 31, 38, 58, 68, 103, 105, 110, 113, 132, 135, 144, 147, 164]

Temporal classes names:

From Kinetics: [bouncing on trampoline, breakdancing, busking, cartwheeling, cleaning shoes, country line dancing, drop kicking, gymnastics tumbling, hammer throw, high kick, jumpstyle dancing, kitesurfing, parasailing, playing cards, playing cymbals, playing drums, playing ice hockey, robot dancing, shining shoes, shuffling cards, side kick, ski jumping, skiing (not slalom or crosscountry), skiing crosscountry, skiing slalom, snowboarding, somersaulting, tap dancing, throwing ball, throwing discus, vault, wrestling]

From Something-something: [Turning something upside down, Approaching something with your camera, Moving something away from the camera, Moving away from something with your camera, Moving something towards the camera, Lifting something with something on it, Moving something away from something, Moving something closer to something, Uncovering something, Pretending to turn something upside down, Covering something with something, Lifting up one end of something, then letting it drop down, Lifting something up completely without letting it drop down, Moving something and something closer to each other, Moving something and something away from each other, Lifting something up completely, then letting it drop down, Stuffing something into something, Moving something and something so they collide with each other]

Static classes names:

From Kinetics: [belly dancing, bending back, blasting sand, blowing nose, changing wheel, clapping, curling hair, deadlifting, dining, doing aerobics, dribbling basketball, eating doughnuts, filling eyebrows, getting a tattoo, laying bricks, long jump, lunge, making bed, moving furniture, mowing lawn, peeling apples, playing badminton, playing controller, playing cricket, pull ups, riding camel, shot put, testifying, trimming trees, waxing eyebrows, yawning, yoga]

From Something-something: [Folding something, Turning the camera upwards while filming something, Holding something next to something, Pouring something into something, Pretending to throw something, Squeezing something, Lifting up one end of something, then letting it drop down, Holding something in front of something, Touching (without moving) part of something, Lifting up one end of something without letting it drop down, Showing something next to something, Poking something so that it falls over, Wiping something off of something, Scooping something up with something, Letting something roll down a slanted surface, Sprinkling something onto something, Pushing something so it spins, Twisting (wringing) something wet until water comes out]

Random sets names:

From Kinetics: [arranging flowers, assembling computer, blowing out candles, bouncing on trampoline, busking, carrying baby, cleaning windows, cooking sausages, curling hair, dancing gangnam style, eating chips, eating doughnuts, egg hunting, feeding goats, gargling, grooming horse, hugging, making snowman, opening bottle, opening present, paragliding, parasailing, passing American football (in game), peeling apples, playing cymbals, playing flute, presenting weather forecast, texting, tossing salad, waiting in line, watering plants, zumba]

From Something-something: [Holding something, Turning the camera left while filming something, Opening something, Moving something up, Showing a photo of something to the camera, Something falling like a feather or paper, Tearing something just a little bit, Pretending to throw something, Attaching something to something, Moving something across a surface without it falling down, Dropping something next to something, Moving something across a surface until it falls down, Pulling something out of something, Pushing something off of something, Pretending or failing to wipe something off of something, Something colliding with something and both are being deflected, Putting something onto something else that cannot support it so it falls down, Pretending to spread air onto something]