Change Log

Change Log#

0.7.0#

Notes#

The first release since 0.6.2 (March 2024). This is both a major structural overhaul of py-feat’s internal Detector class and a coordinated batch of breaking changes that bring py-feat onto modern Python and PyTorch. Users on Python 3.10 or earlier, or who use nltools.data.Adjacency from Fex.distance(), will need to migrate (see the migration guide below).

Highlights:

  • HuggingFace model hub integration: all pre-trained models are now versioned and downloaded on first use.

  • End-to-end MPS support on Apple Silicon (was silently disabled before).

  • Pure-PyTorch head-pose and gaze for the MediaPipe Face Mesh detector. No OpenCV dependency.

  • Modern video decoding via torchcodec (CPU + CUDA; CPU on Apple Silicon, transfer to MPS for inference).

  • Fixed HOGLayer (previously broken; now matches skimage.feature.hog to ~5e-8 absolute tolerance) and wired it into extract_hog_features so the SVM AU and SVM emotion paths actually use the corrected GPU implementation instead of the legacy CPU skimage round-trip.

  • New face_model='retinaface' opt-in: 88.9% WIDERFACE Hard AP (vs img2pose’s 55.5%), batched-by-default with on-device NMS, ~210 detections/sec on Apple Silicon at batch 32.

  • Batched img2pose forward pass: the legacy default detector now runs all frames in a batch through one network call instead of looping per-frame, with vectorized prior generation and on-device postprocessing. ~3-10x faster on multi-frame inputs.

  • Removed nltools, nilearn, av, kornia, and onnx2torch as runtime dependencies.

Planned removals#

  • MPDetector is likely to be removed in a future release. The new Detectorv2 now covers its core functionality (478-point 3D mesh, AUs, emotion, valence/arousal, gaze, head pose, identity) from a single multi-task model and is the recommended path. MPDetector remains available and unchanged in 0.7.

Breaking API changes#

  • Detector.detect_image() and Detector.detect_video() are removed; both flows go through a single Detector.detect() with data_type in {'image', 'tensor', 'video'}.

  • The default face detector is img2pose, which also provides 6-DoF head-pose estimation.

  • Python 3.11+ required. Drops support for 3.8 / 3.9 / 3.10. CI tests 3.11 / 3.12 / 3.13. (PR #262)

  • nltools removed as a dependency. Functions previously sourced from nltools are now in feat.utils.stats: regress, downsample, upsample, set_decomposition_algorithm. Fex.distance() now returns a pandas.DataFrame instead of an nltools.data.Adjacency. (PR #262)

  • nilearn removed as a dependency. Fex.clean() now uses an in-house feat.utils.stats.clean_signal (scipy/numpy-based) for detrending, confound regression, Butterworth filtering, and standardization. Same arguments as before; *args, **kwargs no longer forwarded.

  • av (PyAV) removed as a dependency. VideoDataset now reads frames and metadata via torchcodec only. The metadata['fps_frac'] field (previously a fractions.Fraction from av.stream.average_rate) is gone; use metadata['fps'] (float). (PR #271)

  • AU classifier model files migrated to modern format (*_v2.skops). No user-visible behavior change — predictions are bit-identical to the originals across all 20 sub-classifiers (verified at multiple feature scales and seeds via scripts/migrate_au_skops.py). Internally, the new files (a) reference their wrapper class via the real module path instead of __main__, and (b) for xgb, embed Booster buffers in xgboost’s modern UBJ format. This eliminates a deprecated-API warning and an intermittent SIGSEGV in Booster.__setstate__ on Python 3.13. The original *.skops files remain on HuggingFace alongside the v2 files for compatibility with py-feat ≤ 0.6.x installs; py-feat 0.7+ tries v2 first and falls back to v1 if v2 isn’t yet on the hub.

  • face_model parameter re-introduced on Detector. The v0.7 changelog originally said face detection was hardcoded to img2pose with no swap; that’s no longer accurate. Detector(face_model='img2pose') (default) preserves prior behavior with 6DoF head pose. Detector(face_model='retinaface') activates a new ResNet34-backbone RetinaFace face detector ported from yakhyo’s MIT implementation: 88.9% WIDERFACE Hard AP (vs img2pose’s 55.5% per Cheong et al, Affective Science 2023), batched-by-default with torchvision.ops.batched_nms on-device, ~210 detections/sec on Apple Silicon at batch 32 versus img2pose’s per-image cost. The same face_model='retinaface' kwarg works identically on MPDetector. With this kwarg, the 6DoF pose columns are populated via DLT-PnP from the 68 landmarks (feat.utils.face_pose_pnp); pose values are approximate and can differ from img2pose’s regressed values by up to ~30° on Pitch. Stay on face_model='img2pose' for pose accuracy that matches the published Cheong et al numbers. Trained weights distributed at py-feat/retinaface_r34 on HuggingFace (the repo retains the backbone-tagged name; the user-facing kwarg is just 'retinaface'). Note for v0.7-dev users: an earlier 'retinaface_r34' spelling that worked only on Detector has been dropped — use 'retinaface' everywhere. (PRs #275, #277, #285)

  • mp_facemesh_v2 migrated from pickled FX GraphModule to TorchScript (closes #249). The previous landmark file (face_landmarks_detector_Nx3x256x256_onnx.pth) was an onnx2torch-converted FX graph, which can only be deserialized via torch.load(weights_only=False) — a documented arbitrary-code-execution path — and required onnx2torch to be importable at load time. The new file (face_landmarks_detector.pt, distributed alongside the legacy file at py-feat/mp_facemesh_v2) is a TorchScript module loaded via torch.jit.load: no pickle execution, no onnx2torch import. Bit-identical outputs across batch sizes 1, 2, 4, 8 verified at conversion time. py-feat 0.7+ tries the new file first and falls back to the legacy file (with a warning) for older HF revisions. The legacy face_landmarks_detector_Nx3x256x256_onnx.pth file will be deleted from py-feat/mp_facemesh_v2 shortly after v0.7.0 ships so the unsafe deserialization path stops being a hosted attack surface; py-feat 0.6.x users on fresh installs will need to upgrade to ≥ 0.7.0 once the file is removed (existing 0.6.x installs with the file already cached locally are unaffected).

  • onnx2torch removed as a runtime dependency. Was only needed to deserialize the legacy mp_facemesh_v2 FX GraphModule; with TorchScript loading the package is no longer touched at runtime. Drops one transitive dep + one piece of the security surface.

  • Default identity_model changed from 'facenet' to 'arcface' on both Detector and MPDetector. ArcFace is angular-margin-loss-trained (vs FaceNet’s triplet loss) and produces embeddings that disentangle identity from pose and expression — the dimensions FEAT video data varies along most. On multi_face.jpg, FaceNet’s max off-diagonal cosine similarity between different people was 0.76 (merging identities at typical thresholds); ArcFace’s was 0.35 (clean separation). Inference cost change is +4 % on M5 MBP MPS for the full retinaface_r34 + svm AU + identity pipeline. (MPDetector’s default also moves from None to 'arcface'; users wanting to skip identity computation can pass identity_model=None explicitly.) Users who need to stay on FaceNet for backwards compatibility can opt in explicitly with Detector(identity_model='facenet') or MPDetector(identity_model='facenet'). The ArcFace ResNet50 weights (InsightFace’s buffalo_l w600k_r50.onnx, repackaged as arcface_r50.safetensors at py-feat/arcface_r50 on HuggingFace) are distributed under InsightFace’s non-commercial-research license; the underlying training data (WebFace600K, Tsinghua) is also research-only. For local testing before the HuggingFace artifact lands, set FEAT_ARCFACE_R50_PATH=/path/to/arcface_r50.safetensors. See feat/identity_detectors/arcface/MODEL_CARD.md for the full model card and citation block, and the root LICENSE file for the third-party listing. (PR #284)

New features#

  • Detectorv2: a single multi-task model. A new from feat import Detectorv2 runs one ConvNeXt V2-Tiny network (py-feat/face_multitask_v2) that predicts 20 Action Units, 7-class emotion, continuous valence/arousal, gaze, a 478-point 3D MediaPipe FaceMesh, and 6-DoF head pose in a single forward pass, with RetinaFace detection and an optional ArcFace identity branch. Unlike the modular Detector (v1) — which chains a separate model per task and emits 68-point landmarks — Detectorv2 replaces that chain with one network, making it substantially faster (especially on single frames) and adding valence/arousal + gaze that v1 doesn’t produce. On py-feat’s end-to-end harness it beats the v1 (xgb+resmasknet) detector on AU (DISFA+ 8-AU F1 0.756 vs 0.588) and RAF-DB emotion (0.849 vs 0.542). The model set is fixed (no per-component swapping). See the two-detector overview.

  • au_to_mesh v4 AU → 3D mesh model (new default). The 478-vertex mesh visualization (plot_face_mesh / predict_face_mesh) now defaults to model_version="v4" (py-feat/au_to_mesh), fit from Detectorv2’s own predicted mesh on ~733K CelebV-HQ frames in the 20-AU AU_LANDMARK_MAP['Feat'] space. It is the default mesh model for the MediaPipe-mesh detectors, Detectorv2 and MPDetector; Detector (v1) is unaffected — it emits 68-point landmarks and visualizes AUs with the 68-pt au_to_landmarks model. The neutral mesh is Procrustes-aligned to the canonical MediaPipe face (frontal) and aspect-corrected to natural proportions. v2 (original 20-AU) and v3 (Detectorv2 24-AU) remain selectable via model_version=.

  • Identity detector via facenet. Each detected face is projected into a 512-d embedding space and clustered by cosine similarity; the resulting label is stored in Fex.identities and the embeddings in Fex.identity_embeddings. The clustering threshold defaults to 0.8 and can be tuned at detection time via face_identity_threshold or recomputed on an existing Fex via .compute_identities(threshold=new_threshold).

  • Memory-efficient identity clustering (new default). Fex.compute_identities() / cluster_identities() previously built a dense N×N similarity matrix (connected-components), which OOM’d on long videos (~40 GB at 200K faces). The default is now an incremental “gallery” method (single pass, O(#identities) memory — scales to arbitrarily large inputs) and an "hdbscan" option (density-based, emits "Unknown" for outliers) was added. Behavior change: the gallery default is order-dependent and not transitive, so labels can differ from the old connected-components output; pass method="connected" for the exact legacy behavior. Non-finite (no-detection) embeddings are now labelled NaN. Requires scikit-learn ≥ 1.3 for the hdbscan option.

  • HuggingFace model integration. Pre-trained weights live at the py-feat HuggingFace org and are downloaded on first use.

  • Pure-PyTorch head-pose and gaze for MediaPipe Face Mesh. Closed-form Umeyama similarity alignment replaces the iterative Adam-loop pose estimator; head-pose-compensated gaze in head-centric frame. No OpenCV dependency. (PR #261, wired into MPDetector in #270)

  • feat.utils.io.decode_video for sliced or streamed video decoding via torchcodec; replaces the prior PyAV path and fixes the regression where loading a single frame for plotting decoded the entire video into memory. (PR #263)

  • HOGLayer now produces feature vectors that match skimage.feature.hog to ~5e-8 absolute tolerance for L1, L1-sqrt, L2, and L2-Hys block normalizations (PR #259), and extract_hog_features now uses HOGLayer end-to-end so the SVM AU and SVM emotion paths inherit the speedup and the corrected feature output instead of round-tripping through skimage on CPU (PR #266).

  • Batched img2pose forward pass. The default face/pose detector now runs an entire batch through a single network call instead of looping per frame, with vectorized prior generation and on-device postprocessing. Combined with the MPS device handoff in #258, multi-frame Detector(face_model='img2pose').detect() calls are ~3-10× faster on Apple Silicon and CUDA. (PR #274)

  • Vectorized per-frame post-processing. invert_padding_to_results replaces the prior per-frame loop with a single set of vectorized numpy ops that fold the dataloader’s padding/scaling back into the output bboxes and landmarks; ~25× faster on long videos with many landmarks. (PR #267)

  • MPDetector surfaces 20 predicted FACS AUs alongside its 52 MediaPipe blendshapes. A new linear PLS regressor (bs_to_au_pls_v2, py-feat/bs_to_au) trained on ~350K paired (xgb AU, MP blendshape) frames from 10K CelebV-HQ celebrity videos translates blendshapes → AUs at inference. Closes a long-standing schema mismatch where MPDetector’s Fex declared au_columns but the actual DataFrame contained the 52 blendshapes. On a DISFA+ benchmark (n=57,150 frames, 12 labeled AUs), the new path beats the standard Detector xgb-on-dlib baseline on mean Pearson r (0.51 vs 0.40). (PRs #300, #302)

  • Refreshed default AU → 68-pt landmark visualization model. feat.plotting.plot_face and feat.plotting.load_viz_model now default to a new PLSAULandmarkModel (v2 PLS, py-feat/au_to_landmarks) trained on ~350K paired (AU, landmark) frames from ~10K CelebV-HQ celebrity videos with pose covariates and pose×AU interactions absorbed at training time, dropping out at inference. OOS R² = 0.794 ± 0.004 (3-fold GroupKFold by video) vs ~0.5 for the legacy v1 model. The legacy pyfeat_aus_to_landmarks filename still loads the original PLSRegression from the same HF repo for backwards compatibility. (PR #301)

  • Unified dlib-68 landmark accessor. Fex.landmarks_dlib68_xy(row=None) returns 68-point landmarks in dlib layout regardless of whether the Fex came from Detector (already 136-d dlib) or MPDetector (subsamples 68 dlib-equivalent vertices from the 478-vertex MP mesh via the DLIB68_FROM_MP478 index map). Fex.plot_detections(faces='landmarks') now works on MPDetector output (was silently broken before). (PR #303)

  • 3D MediaPipe FaceMesh visualization from AU intensities. New opt-in feat.plotting.plot_face_mesh(au=...) and predict_face_mesh(au) consume a PLS regressor (au_to_mesh_pls_v2, py-feat/au_to_mesh) that maps 20 AU intensities + 3 pose covariates to a 478-vertex 3D mesh in a pose-canonical Procrustes-aligned frame. Renders as a matplotlib 3D wireframe over the MediaPipe canonical contours (lips + eyes + eyebrows + face oval). (PR #304)

  • 68-pt dlib → 478-pt MediaPipe mesh bridge. New predict_mesh_from_dlib68(landmarks) reconstructs a 478-vertex 3D mesh from raw Detector landmarks via a PCA-bottleneck linear regression (landmarks68_to_mesh478_pca_v2, py-feat/landmarks68_to_mesh478) trained on ~340K paired frames; OOS R² ≈ 0.48. Lets users with a Detector Fex render the 3D mesh through plot_face_mesh(mesh=...) without running MPDetector. (PR #305)

  • Shared 2D Procrustes alignment helper. feat.utils.image_operations.procrustes_align_2d_batched(coords, anchor_idx, ref_anchors) performs batched 2D Umeyama similarity alignment of a (n, K, 2) landmark batch to a reference frame defined by anchor indices. Hoisted out of the bridge implementation so other landmark-canonicalization paths can reuse it. (PR #306)

  • Interactive 3D Plotly backend for face mesh viz. feat.plotting.plot_face_mesh_plotly() returns a plotly.graph_objects.Figure with the predicted mesh rendered as a single Scatter3d lines trace (NaN-separated edges for fast rendering). Two modes: 'tesselation' (default, full 2,556-edge MP tessellation, dense 3D look) or 'contours' (124-edge canonical contours, matches the matplotlib version). Same au= / mesh= source dispatch as plot_face_mesh. Plotly is already a runtime dependency; no new install. (PR #307)

Bug fixes#

  • MPS device handoff. Three places in the inference path created tensors on CPU while model weights were on MPS or CUDA, causing Detector(device='mps') to crash partway through. End-to-end MPS detection now runs cleanly on Apple Silicon. (PR #258)

  • Detector / MPDetector correctness fixes. Several pre-existing bugs in the multi-face / batched paths produced wrong landmarks, mismatched dtypes, or wrong frame-id assignments depending on input shape. End-to-end Fex outputs now match expected values across (image, video) × (single-face, multi-face) × (batch=1, batch>1) combinations. (PR #264)

  • MPDetector(face_model='retinaface') works again. The v0.7 RetinaFace rebuild deleted the MobileNet0.25 path that MPDetector’s face detector relied on; the prior workaround was a NotImplementedError. MPDetector now uses the same ResNet34 wrapper as Detector, so MPDetector(face_model='retinaface') and MPDetector(face_model='retinaface_r34') both build and detect end-to-end. (PR #278)

  • MPDetector pose dtype mismatch fixed. A pre-existing bug in convert_landmarks_3d produced np.float64 landmarks (via astype(float)) while MediaPipe’s canonical face model is float32. estimate_face_pose_from_mesh’s matmul would have raised expected m1 and m2 to have the same dtype. Forced astype("float32") to fix. (PR #278)

  • MPDetector pose Y/Z-axis convention mismatch fixed. MediaPipe Face Mesh outputs landmarks in image-pixel space (Y down, Z into screen); the canonical face model lives in head-centric space (Y up, Z out of face). The Umeyama alignment was absorbing the convention difference into a 180° rotation about X, which surfaced as Pitch clustering near ±π for every forward-facing portrait. convert_landmarks_3d now flips Y and Z at the conversion boundary so the alignment recovers the actual head pose. After the fix, forward-facing portraits return |Pitch| < 30° instead of ~180°. (PR #279)

  • Detector(device='auto') / MPDetector(device='auto') crash fixed. When loading the emotion / identity / AU sub-models, torch.load was called with map_location=device (the raw constructor argument) instead of the resolved self.device. With device='auto' this passed the literal string "auto" to torch.load, raising RuntimeError: don't know how to restore data location ... tagged with auto — on the exact device='auto' path the install docs recommend. All five load sites now use self.device (which set_torch_device resolves to cuda/mps/cpu). Detectorv2 was unaffected.

  • iplot_detections crash on Detectorv2 detections fixed. Fex.plot_singleframe_detections (used by iplot_detections) hardcoded the v1 emotion names ["anger", "disgust", "fear", "happiness", "sadness", "surprise", "neutral"] when building each face’s emotion annotation, so it raised KeyError on Detectorv2’s capitalized columns (Neutral/Happy/…) — even with emotions=False, since the annotation is built unconditionally. Now reads row[self.emotion_columns], so interactive plots work for both detectors.

Documentation updates#

  • Tutorials updated for the new Detector.detect() API.

  • New FAQ.

Migration guide#

If you used:

  • Detector.detect_image() or Detector.detect_video() -> change to Detector.detect(..., data_type='image') or data_type='video'.

  • from nltools.stats import regress, downsample, upsample -> change to from feat.utils.stats import regress, downsample, upsample.

  • from nltools.data import Adjacency for the result of Fex.distance() -> the result is now a plain pandas.DataFrame. Drop the import; use .shape instead of .square_shape().

  • Python 3.10 or earlier -> upgrade to 3.11+. macOS users may also need brew install libomp for xgboost on Python 3.13.

MPDetector pose / gaze numerics changed#

If you compared Pitch / Roll / Yaw / X / Y / Z outputs from MPDetector between 0.6.x and 0.7.0, the values will differ. The prior estimator minimized the wrong objective (mean(z_proj²) instead of a true reprojection error) and produced effectively meaningless head-pose values; the new closed-form Umeyama alignment produces the actual head pose. Treat the prior values as noise.

Pose translation columns now in original-frame pixels. The Umeyama alignment that recovers head pose is scale-invariant for rotation (Pitch/Roll/Yaw unchanged), but the translation term (X/Y/Z columns of the 6DoF pose) moves with the input scale. As of the perf refactor that folds DataLoader Rescale-inversion into forward() (replacing the post-hoc invert_padding_to_results step), MPDetector reads landmarks from the Fex in original-frame pixel coords rather than padded-frame coords. So X/Y/Z magnitudes in Fex.facepose are now in the same coordinate system as FaceRectX, x_0, etc. — typically smaller than before and shifted, but in a more useful frame.

MPDetector also now emits gaze_pitch and gaze_yaw columns (radians, head-centric frame) in addition to the existing gaze_angle. The gaze_angle column is preserved for backward compatibility but its semantic shifted from “angle from camera-forward” to “angle from head-forward in head frame” - so a turned head no longer registers as averted gaze even when the eyes are looking along the head’s forward axis.

0.6.1#

Notes#

This version drops support for Python 3.7 and fixes several dependency related issues:

  • #162

  • #176

  • We can now handle images with an alpha-channel by just grabbing the RGB channels (typically in png files)

  • Update minimum scikit-learn version requirement to ensure our viz models are loaded correctly

  • Soft-pin numexpr version until this upstream pandas issue is fixed

0.6.0#

Notes#

This is a large model-update release. Several users noted issues with our AU models due to problematic HOG feature extraction. We have now retrained all of our models that were affected by this issue. This version will automatically download the new model weights and use them without any additional user input.

Detector Changes#

We have made the decision to make video processing much more memory efficient at the trade-off of increased processing time. Previously py-feat would load all frames into RAM and then process them. This was problematic for large videos and would cause kernel panics or system freezes. Now, py-feat will lazy-load video-frames one at a time, which scales to videos of any length or size assuming that your system has enough RAM to hold a few frames in memory (determined by batch_size). However, this also makes processing videos a bit slower and GPU benefits less dramatic. We have made this trade-off in favor of an easier end-user experience, but will be watching torch’s VideoReader implementation closely and likely use that in future versions.

0.5.1#

Notes#

This is a maintenance release that addresses multiple under-the-hood issues with py-feat failing when images or videos contain 0 faces. It addresses the following specific issues amongst others and is recommended for all users:

0.5.0#

Notes#

This is a large overhaul and refactor of some of the core testing and API functionality to make future development, maintenance, and testing easier. Notable highlights include:

  • tighter integration with torch data loaders

  • dropping opencv as a dependency

  • experimental support for macOS m1 GPUs

  • passing keyword arguments to underlying torch models for more control

Detector Changes#

New#

  • you can now pass keyword arguments directly to the underlying pytorch/sklearn models on Detector initialization using dictionaries. For example you can do: detector = Detector(facepose_model_kwargs={'keep_top_k': 500}) to initialize img2pose to only use 500 instead of 750 features

  • all .detect_* methods can also pass keyword arguments to the underlying pytorch/sklearn models, albeit these will be passed to their underlying __call__ methods

  • SVM AU model has been retrained with new HOG feature PCA pipeline

  • new XGBoost AU model with new HOG feature PCA pipeline

  • .detect_image and .detect_video now display a tqdm progressbar

  • new skip_failed_detections keyword argument to still generate a Fex object when processing multiple images and one or more detections fail

Breaking#

  • the new default model for landmark detection was changed from mobilenet to mobilefacenet.

  • the new default model for AU detection was changed to our new xgb model which gives continuous valued predictions between 0-1

  • remove support for fer emotion model

  • remove support for jaanet AU model

  • remove support for logistic AU model

  • remove support for pnp facepose detector

  • drop support for reading and manipulating Affectiva and FACET data

  • .detect_image will no longer resize images on load as the new default for output_size=None. If you want to process images with batch_size > 1 and images differ in size, then you will be required to manually set output_size otherwise py-feat will raise a helpful error message

Fex Changes#

New#

  • new .update_sessions() method that returns a copy of a Fex frame with the .sessions attribute updated, making it easy to chain operations

  • .predict() and .regress() now support passing attributes to X and or Y using string names that match the attribute names:

    • 'emotions' use all emotion columns (i.e. fex.emotions)

    • 'aus' use all AU columns (i.e. fex.aus)

    • 'poses' use all pose columns (i.e. fex.poses)

    • 'landmarks' use all landmark columns (i.e. fex.landmarks)

    • 'faceboxes' use all facebox columns (i.e. fex.faceboxes)

    • You can also combine feature groups using a comma-separated string e.g. fex.regress(X='emotions,poses', y='landmarks')

  • .extract_* methods now include std and sem. These are also included in .extract_summary()

Breaking#

  • All Fex attributes have been pluralized as indicated below. For the time-being old attribute access will continue to work but will show a warning. We plan to formally drop support in a few versions

    • .landmark -> .landmarks

    • .facepose -> .poses

    • .input -> .inputs

    • .landmark_x -> .landmarks_x

    • .landmark_y -> .landmarks_y

    • .facebox -> .faceboxes

Development changes#

  • test_pretrained_models.py is now more organized using pytest classes

  • added tests for img2pose models

  • added more robust testing for the interaction between batch_size and output_size

General Fixes#

  • data loading with multiple images of potentially different sizes should be faster and more reliable

  • fix bug in resmasknet that would give poor predictions when multiple faces were present and particularly small

  • #150

  • #149

  • #148

  • #147

  • #145

  • #137

  • #134

  • #132

  • #131

  • #130

  • #129

  • #127

  • #121

  • #104

0.4.0#

Major version breaking release!#

  • This release includes numerous bug fixes, api updates, and code base changes make it largely incompatible with previous releases

  • To fork development from an older version of py-feat you can use this archival repo instead

New#

  • Added animate_face and plot_face functions in feat.plotting module

  • Fex data-classes returned from Detector.detect_image() or Detector.detect_video() now store the names of the different detectors used as attributes: .face_model, .au_model, etc

  • The AU visualization model used by plot_face and Detector.plot_detections(faces='aus') has been updated to include AU11 and remove AU18 making it consistent with Py-feat’s custom AU detectors (svm and logistic)

  • A new AU visualization model supporting the jaanet AU detector, which only has 12 AUs, has now been added and will automatically be used if Detector(au_model='jaanet').

    • This visualization model can also be used by the plot_face function by by passing it to the model argument: plot_face(model='jaanet_aus_to_landmarks')

Breaking Changes#

  • Detector no longer support uninitialized models, e.g. any_model = None

    • This is is also true for Detector.change_model

  • Columns of interest on Fex data classes were previously accessed like class methods, i.e. fex.aus(). These have now been changed to class attributes, i.e. fex.aus

  • Remove support for DRML AU detector

  • Remove support for RF AU and emotion detectors

  • New default detectors:

    • svm for AUs

    • resmasknet for emotions

    • img2pose for head-pose

Development changes#

  • Revamped pre-trained detector handling in new feat.pretrained module

  • More tests including testing all detector combinations

Fixes#

0.3.7#

  • Fix import error due to missing init

0.3.6#

  • Trigger Zenodo release

0.2.0#

  • Testing pypi upload

Contents