There has been significant progress on pose estimation and tracking in recent years. Yet, much of the focus has been on developing complex solutions involving multi-component systems, cascaded networks, and sophisticated loss functions. In this work, we ask a simple question: how good could a simple method be? Specifically, we provide simple baseline methods for both pose estimation and pose tracking.

For pose estimation, our method involves a ResNet backbone augmented with a few deconvolutional layers to generate heatmaps for joint detection. For pose tracking, our method involves assigning poses based on greedy matching of detected keypoints across frames using optical flow. Despite the simplicity of these methods, they achieve competitive or even superior results compared to state-of-the-art methods on challenging benchmarks such as COCO and PoseTrack.

Human pose estimation, the task of determining the spatial locations of body joints or keypoints from an image, has attracted significant attention due to its widespread applications in action recognition, human-computer interaction, and augmented reality. Recent approaches to the problem have achieved impressive results by employing increasingly complex network architectures and training strategies. However, the growing complexity makes it difficult to identify which design choices are truly essential for achieving high performance.

We believe that establishing strong and simple baselines is important for understanding the core problem and for measuring progress in the field. Our goal is not to propose a new state-of-the-art method but rather to show that a straightforward approach, when implemented carefully with proper training procedures, can achieve surprisingly strong performance. The key insight is that a simple network architecture combined with proper training details can match or exceed the performance of more complex designs.

Our pose estimation network adopts the structure of adding a few deconvolutional layers over the last convolution stage in the ResNet architecture. Three deconvolutional layers with batch normalization and ReLU activation are used. Each layer has 256 filters with 4x4 kernel. A 1x1 convolutional layer is added at last to generate predicted heatmaps for all k key points. The overall architecture is remarkably simple and it can be generated by the ResNet backbone and a few transposed convolution layers with no bells and whistles.

For multi-person pose tracking in videos, we adopt a simple approach based on optical flow. Given detected poses in a current frame, we propagate each keypoint to the next frame using optical flow and perform greedy bipartite matching between the propagated keypoints and the newly detected keypoints. The matching cost is computed as the Object Keypoint Similarity (OKS) between the propagated and detected poses. This simple strategy avoids the need for complex tracking algorithms such as graph-based optimization or recurrent networks.

We evaluate our approach on the COCO keypoint detection benchmark. Our models are trained on the COCO train2017 dataset including 57K images and 150K person instances. We report results on the val2017 set with 5000 images and test-dev2017 set. The evaluation metric is Object Keypoint Similarity (OKS) based AP. We use the person detector from the two-stage top-down pipeline and apply our pose estimation network on each detected person bounding box.

With ResNet-152 as the backbone, our simple baseline achieves 73.7 AP on the COCO test-dev set, which is competitive with the state-of-the-art results from much more complex methods such as CPN which uses a cascaded network architecture with online hard keypoint mining. When using the same person detector and ResNet-50 backbone, our method achieves 70.4 AP, which already outperforms many existing approaches. The results clearly demonstrate that most of the improvement in recent methods comes from better backbone networks and training strategies rather than architectural innovations.

For pose tracking on the PoseTrack dataset, our simple baseline with greedy matching achieves competitive tracking accuracy compared to methods using more sophisticated temporal modeling. Specifically, our approach obtains multi-person pose tracking accuracy of 65.4 MOTA on the PoseTrack validation set. While more complex temporal models may offer marginal improvements, the gap is surprisingly small, suggesting that the pose estimation quality is the primary bottleneck for tracking performance.

We have provided simple yet strong baseline methods for human pose estimation and tracking. For pose estimation, we show that a ResNet backbone with a few deconvolutional layers can achieve surprisingly strong results. For pose tracking, a simple greedy matching strategy based on optical flow is sufficient to achieve competitive performance. Our results suggest that the community should pay more attention to training details and baseline comparisons rather than solely focusing on architectural novelty.

We hope that our work can serve as a reference point for future research, providing a clear and reproducible baseline against which new methods can be evaluated. We acknowledge that there is certainly room for improvement through more sophisticated approaches, but we argue that any new method should demonstrate significant advantages over these simple baselines to justify its added complexity.