Figure 7.
Benchmark comparison between the existing standards and LabGym
(A) The workflow of the benchmark comparison between the combination of DeepLabCut and B-SOiD (DLC + B-SOiD) and LabGym, which was performed on the same computer that has a 3.6 GHz Intel Core i9-9900K CPU, a 64 GB memory, and an NVIDIA GeForce RTX 2080 GPU. The setting of Categorizer in LabGym: Animation Analyzer: level 1 with input shape 8 × 8 × 1 × 20; Pattern Recognizer: level 2 with input shape 16 × 16 × 3.
(B) The total time spent for DLC + B-SOiD and LabGym to finish the workflow in (A).
(C and D) Comparisons of frame-wise behavior classifications between the computer software and the ground truth (expert human annotation) in two testing videos: baseline (C) and post treatment (D). The raster plots show frame-wise behavior events with the x axis indicating the time (1 min). Different colors indicate different behavior types.
(E) Comparison of frame-wise behavior classifications between the computer software and the ground truth (expert human annotation) in generalization to a different testing enclosure. The raster plots show frame-wise behavior events with the x axis indicating the time (1 min). The colors match those for behaviors shown in (C).