Estimating multi-person poses using greedy part assignment

1. An apparatus for estimating poses, comprising a processor to:

estimate a number of people based on a significant head count of a plurality of received refined part detections refined based on unary probabilities of part detections exceeding a threshold unary probability;

cluster the refined part detections based on the estimated number of people to generate clustered part detections;

select a number of candidate person clusters for each clustered part detection based on proximity to the clustered part detection;

calculate a cluster affinity score based on pairwise probabilities for each combination of a predecessor part of the candidate person cluster and clustered part detection, and greedily sequentially assign each clustered part detection to a candidate person cluster based on the cluster affinity score to generate person clusters; and

generate a pose for each person cluster.

2. The apparatus of claim 1 , wherein the processor is to detect and suppress hallucinated parts in the clustered part detections, wherein clustered part detections having a structural compliance probability score below a threshold score are to be detected as hallucinated parts.

3. The apparatus of claim 1 , wherein the significant head count is based on a unary probability for each detected head exceeding a threshold probability.

4. The apparatus of claim 1 , wherein the processor is to cluster the refined part detections based on a spatial k-means clustering, wherein a refined part detection that is closer to a center of a generated spatial cluster is to be selected as a clustered part detection.

5. The apparatus of claim 1 , wherein the cluster affinity score calculated for each combination is based on a received pairwise probability between the clustered part detection and one or more predecessor parts of the candidate person cluster.

6. The apparatus of claim 1 , wherein the cluster affinity score calculated for each combination is based on a received pairwise probability between the clustered part detection and one or more predecessor parts on a same side of the candidate person cluster.

7. The apparatus of claim 1 , wherein the processor is to process refined part detections of a part class in parallel.

8. The apparatus of claim 1 , wherein the processor is to process clustered part detections of an identical part class in parallel.

9. The apparatus of claim 1 , wherein the refined part detections comprise part detections with unary probabilities exceeding a threshold probability.

10. The apparatus of claim 1 , comprising a neural network trained to detect the body part candidates.

11. A method for estimating multi-person poses, comprising:

estimating, via the processor, a number of people based on a significant head count of a plurality of received refined part detections refined based on unary probabilities of part detections exceeding a threshold unary probability;

clustering, via the processor, the refined part detections based on the estimated number of people to generate clustered part detections;

selecting, via the processor, a number of candidate person clusters for each clustered part detection based on proximity to the clustered part detection;

calculating, via the processor, a cluster affinity score based on pairwise probabilities for each combination of a predecessor part of the candidate person cluster and clustered part detection, and greedily sequentially assigning, via the processor, each clustered part detection to a candidate person cluster based on the cluster affinity score to generate person clusters; and

generating, via the processor, a pose for each person cluster.

12. The method of claim 11 , wherein selecting the number of candidate person clusters and calculating the cluster affinity score is performed for clustered part detections of the same part class in parallel.

13. The method of claim 11 , wherein clustering the refined part detections comprises performing a spatial k-means clustering.

14. The method of claim 11 , comprising detecting and suppressing, via the processor, hallucinated parts in the clustered part detections.

15. The method of claim 11 , wherein calculating a cluster affinity score comprises calculating co-occurrence probabilities for the combinations and greedily sequentially assigning the clustered part detections comprises assigning the clustered part detections by type of body part from head down.

16. At least one computer readable medium for estimating multi-person poses having instructions stored therein that, in response to being executed on a computing device, cause the computing device to:

receive body part candidates and

refine the body part candidates based on unary probabilities of part detections exceeding a threshold unary probability to generate refined part detections;

estimate a number of people based on a significant head count of the refined part detections;

cluster the refined part detections based on the estimated number of people to generate clustered part detections;

select a number of candidate person clusters for each clustered part detection based on proximity to the clustered part detection;

calculate a cluster affinity score based on pairwise probabilities for each combination of a predecessor part of the candidate person cluster and clustered part detection;

greedily sequentially assign each clustered part detection to a candidate person cluster based on the cluster affinity score to generate person clusters; and

generate a pose for each person cluster.

17. The at least one computer readable medium of claim 16 , comprising instructions to process clustered part detections of a particular part type in parallel.

18. The at least one computer readable medium of claim 16 , comprising instructions to detect a significant head based on a unary probability for a detected head exceeding a threshold probability.

19. The at least one computer readable medium of claim 16 , comprising instructions to detect and suppress hallucinated parts based on a structural compliance probability score.

20. The at least one computer readable medium of claim 16 , comprising instructions to cluster the refined part detections based on a spatial k-means clustering.