qhjqhj00 activity-recognition-eval
Simultaneous Learning from Human Pose and Object Cues for Real-Time Activity Recognition — Reily et al. (2020) (arXiv:2004.03453, 2020)
What this evaluates
Evaluates a model's ability to recognize human activities in real-time by simultaneously learning from skeletal pose data and object attributes. It probes the integration of multi-modal cues (color, shape, distance, or object probabilities) for accurate and efficient activity classification in robotics scenarios.
Datasets
- Cornell Activity Dataset (CAD-60) — total ?; splits: test (-1)
- MSR Daily Activity 3D Dataset — total ?; splits: test (-1)
Metrics
accuracy(primary) — range: percent- Percentage of correctly identified activity executions out of the total number of executions.
Processing Speed (Hz)— range: other- Number of recognition processing cycles per second, calculated as 1 divided by time per frame.
Input / output format
Input: Per instance: color and depth images, annotated 3D skeleton joint positions (15 joints for CAD-60, 20 for MSR), and object attributes (RGB histograms, HOG features, 3D object-joint distances, or YOLO-derived object category probabilities).
Output: Activity class label (classification) and a weight matrix U indicating the importance of each object attribute to the predicted activity.
Scoring recipe
def compute_accuracy(predictions, gold_labels):
correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
return (correct / len(gold_labels)) * 100
def compute_speed(time_per_frame_sec):
return 1.0 / time_per_frame_sec
Common pitfalls
- Dataset train/val/test splits are not explicitly defined in the paper.
- Object attributes must be extracted using specific modalities (RGB histograms, HOG, or YOLO probabilities) rather than raw images.
- Real-time speed benchmarks are hardware-dependent (tested on Intel i5 with 4GB RAM).
Evidence (verbatim from paper)
It is observed that our approach achieves an activity recognition accuracy of 97.71%.
Citation
@misc{reily2020simultaneous,
title={Simultaneous Learning from Human Pose and Object Cues for Real-Time Activity Recognition},
author={Reily et al. (2020)},
year={2020},
note={arXiv:2004.03453}
}
- arXiv: 2004.03453