Datasets:

py-feat
/

benchmarks

tool stringclasses 6 values	dataset stringclasses 10 values	modality stringclasses 4 values	metric stringlengths 7 16	value float64 0 44.6	n float64 800 57.7k ⌀	train_status stringclasses 4 values	notes stringclasses 13 values
pyfeat_v2	disfaplus	au	f1_mean	0.5396	57,150	out_of_sample	truth>=2, prob>=0.5; 12 AU
pyfeat_v2	disfaplus	au	f1_AU01	0.7837	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU02	0.5674	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU04	0.7079	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU05	0.5788	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU06	0.5858	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU09	0.8252	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU12	0.559	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU15	0.649	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU17	0.2816	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU20	0.0119	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU25	0.6682	57,150	out_of_sample	null
pyfeat_v2	disfaplus	au	f1_AU26	0.2563	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_mean	0.2503	57,150	out_of_sample	truth>=2, prob>=0.5; 12 AU
pyfeat_v1	disfaplus	au	f1_AU01	0.3393	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU02	0.2361	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU04	0.429	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU05	0	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU06	0.3213	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU09	0.1334	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU12	0.3927	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU15	0.1179	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU17	0.1368	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU20	0.096	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU25	0.6813	57,150	out_of_sample	null
pyfeat_v1	disfaplus	au	f1_AU26	0.1202	57,150	out_of_sample	null
openface3	disfaplus	au	f1_mean	0.4882	57,668	out_of_sample	8 AU predicted; missing AU treated as 0 over 12
openface3	disfaplus	au	f1_AU01	0.8476	57,668	out_of_sample	null
openface3	disfaplus	au	f1_AU02	0.6899	57,668	out_of_sample	null
openface3	disfaplus	au	f1_AU04	0.8354	57,668	out_of_sample	null
openface3	disfaplus	au	f1_AU06	0.4202	57,668	out_of_sample	null
openface3	disfaplus	au	f1_AU09	0.8016	57,668	out_of_sample	null
openface3	disfaplus	au	f1_AU12	0.7375	57,668	out_of_sample	null
openface3	disfaplus	au	f1_AU25	0.9338	57,668	out_of_sample	null
openface3	disfaplus	au	f1_AU26	0.5919	57,668	out_of_sample	null
libreface	disfaplus	au	f1_mean	0.4614	57,150	out_of_sample	research RepVGG; truth>=2, intensity>=2; 12 AU
libreface	disfaplus	au	f1_AU01	0.7031	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU02	0.7057	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU04	0.7462	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU05	0.6256	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU06	0.1116	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU09	0.7505	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU12	0.5632	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU15	0.0671	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU17	0.0648	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU20	0.0061	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU25	0.9065	57,150	out_of_sample	null
libreface	disfaplus	au	f1_AU26	0.2869	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU01	0.7296	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU02	0.7157	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU04	0.8852	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU05	0.8449	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU06	0.7509	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU09	0.8643	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU12	0.8451	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU15	0.5913	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU17	0.6065	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU20	0.2439	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU25	0.9421	57,150	out_of_sample	null
libreface	disfaplus	au	pcc_AU26	0.7767	57,150	out_of_sample	null
pyafar	disfaplus	au	f1_mean_7au	0.2597	57,150	out_of_sample	7 of 12 AU overlap; occ>=0.5 vs truth>=2
pyafar	disfaplus	au	f1_AU01	0.0031	57,150	out_of_sample	null
pyafar	disfaplus	au	f1_AU02	0.398	57,150	out_of_sample	null
pyafar	disfaplus	au	f1_AU04	0	57,150	out_of_sample	null
pyafar	disfaplus	au	f1_AU06	0.6067	57,150	out_of_sample	null
pyafar	disfaplus	au	f1_AU12	0.5483	57,150	out_of_sample	null
pyafar	disfaplus	au	f1_AU15	0.1257	57,150	out_of_sample	null
pyafar	disfaplus	au	f1_AU17	0.1364	57,150	out_of_sample	null
pyfeat_v2	affectnet_val	emotion	accuracy	0.4919	992	held_out	null
pyfeat_v2	affectnet_val	emotion	f1_macro	0.4787	992	held_out	null
pyfeat_v2	affectnet_val	valence_arousal	valence_ccc	0.5349	992	held_out	null
pyfeat_v2	affectnet_val	valence_arousal	arousal_ccc	0.4822	992	held_out	null
pyfeat_v2	raf_test	emotion	accuracy	0.6555	3,019	held_out	null
pyfeat_v2	raf_test	emotion	f1_macro	0.5275	3,019	held_out	null
pyfeat_v2	columbia_gaze	gaze	angular_mae_deg	2.721	1,176	in_sample	a=pitch,b=yaw;sign_pitch=1;sign_yaw=-1;unit=rad
pyfeat_v2	ethxgaze	gaze	angular_mae_deg	44.566	2,763	out_of_sample	a=pitch,b=yaw;sign_pitch=-1;sign_yaw=1;unit=rad
pyfeat_v2	eyediap	gaze	angular_mae_deg	19.975	800	out_of_sample	a=pitch,b=yaw;sign_pitch=-1;sign_yaw=1;unit=rad
pyfeat_v2	eyediap_norm	gaze	angular_mae_deg	19.418	800	out_of_sample	a=pitch,b=yaw;sign_pitch=-1;sign_yaw=1;unit=rad
pyfeat_v2	gaze360	gaze	angular_mae_deg	9.416	2,425	held_out	a=pitch,b=yaw;sign_pitch=1;sign_yaw=1;unit=rad
pyfeat_v2	mpiifacegaze	gaze	angular_mae_deg	2.802	3,000	in_sample	a=pitch,b=yaw;sign_pitch=1;sign_yaw=-1;unit=rad
pyfeat_v2	mpiigaze	gaze	reported_mae_deg	3.92	null	unknown	au_deep v2.4 final (shipped ckpt v24fix_s3; self-reported, in-distribution)
pyfeat_v2	gaze360	gaze	reported_mae_deg	6.81	null	held_out	au_deep v2.4 final (shipped ckpt v24fix_s3; self-reported, in-distribution)
openface3	mpiifacegaze	gaze	reported_mae_deg	2.56	null	in_sample	OpenFace 3.0 paper (arXiv:2506.02891)
openface3	gaze360	gaze	reported_mae_deg	10.6	null	in_sample	OpenFace 3.0 paper (arXiv:2506.02891)
l2cs_net	mpiigaze	gaze	reported_mae_deg	3.92	null	unknown	L2CS-Net paper (arXiv:2203.03339); py-feat's gaze lineage
l2cs_net	gaze360	gaze	reported_mae_deg	10.41	null	unknown	L2CS-Net paper (arXiv:2203.03339); py-feat's gaze lineage
openface3	affectnet_val	emotion	accuracy	0.4929	980	held_out	null
openface3	affectnet_val	emotion	f1_macro	0.5198	980	held_out	null
openface3	raf_test	emotion	accuracy	0.5132	2,233	out_of_sample	null
openface3	raf_test	emotion	f1_macro	0.469	2,233	out_of_sample	null
openface3	columbia_gaze	gaze	angular_mae_deg	12.049	1,176	out_of_sample	a=pitch,b=yaw;sign_pitch=1;sign_yaw=1;unit=deg
openface3	ethxgaze	gaze	angular_mae_deg	37.702	2,477	out_of_sample	a=pitch,b=yaw;sign_pitch=1;sign_yaw=1;unit=rad
openface3	eyediap	gaze	angular_mae_deg	21.257	800	out_of_sample	a=pitch,b=yaw;sign_pitch=1;sign_yaw=1;unit=rad
openface3	eyediap_norm	gaze	angular_mae_deg	21.533	800	out_of_sample	a=pitch,b=yaw;sign_pitch=1;sign_yaw=1;unit=rad
openface3	gaze360	gaze	angular_mae_deg	41.091	2,397	in_sample	a=pitch,b=yaw;sign_pitch=-1;sign_yaw=1;unit=rad
openface3	mpiifacegaze	gaze	angular_mae_deg	7.027	2,879	in_sample	a=pitch,b=yaw;sign_pitch=1;sign_yaw=1;unit=rad
libreface	affectnet_val	emotion	accuracy	0.4553	984	held_out	null
libreface	affectnet_val	emotion	f1_macro	0.4026	984	held_out	null
libreface	raf_test	emotion	accuracy	0.6458	2,705	unknown	null
libreface	raf_test	emotion	f1_macro	0.3855	2,705	unknown	null

End of preview. Expand in Data Studio

py-feat benchmarks

Live benchmark data for py-feat and a cross-tool comparison against OpenFace 3.0, LibreFace, and PyAFAR. Powers the py-feat live dashboard. Updated by scheduled benchmark runs.

Files

File	What
`accuracy.csv`	Tidy long table: one row per `(tool, dataset, modality, metric)`. Covers AU F1 (DISFA+), 7-class emotion (AffectNet-val, RAF-DB), valence/arousal CCC (AffectNet-val), and gaze angular error (Columbia).
`throughput.csv`	py-feat detector frames/sec across hardware × batch.
`cross_tool_methodology.md`	How every tool was run end-to-end, the feature matrix, dataset/protocol choices, and per-tool integration notes.
`competitors/*.json`	Raw per-tool result JSONs (provenance for every number in `accuracy.csv`).

`accuracy.csv` columns

tool · dataset · modality (au / emotion / valence_arousal / gaze) · metric · value · n (samples scored) · notes.

Protocol in one paragraph

Every competitor runs end-to-end as shipped (its own face detector, its own models) on identical images/labels frozen into shared manifests — never as a py-feat dependency, each in its own isolated env. Held-out sets only (DISFA+, not DISFA, since DISFA trains several competitors). Emotion is scored on the 7 shared classes; gaze I/O conventions are resolved identically and in every tool's favor. See cross_tool_methodology.md for the full story, including what it took to get each competitor to run.

Downloads last month: -

py-feat benchmarks

Files

accuracy.csv columns

Protocol in one paragraph

`accuracy.csv` columns