--- dataset_info: features: - name: id dtype: string - name: image dtype: image - name: height dtype: int16 - name: weight dtype: int16 - name: body_type dtype: string - name: shirt_size dtype: string - name: pant_size dtype: string - name: body_pose_params list: float64 - name: pred_global_rots list: list: list: float64 - name: focal_length dtype: float64 - name: pred_joint_coords list: list: float64 - name: global_rot list: float64 - name: shape_params list: float64 - name: pred_cam_t list: float64 - name: pred_keypoints_2d list: list: float64 - name: pred_keypoints_3d list: list: float64 - name: bbox list: float64 - name: scale_params list: float64 - name: bmi dtype: float64 - name: height_cls dtype: class_label: names: '0': H_132_155 '1': H_157_157 '2': H_160_160 '3': H_163_163 '4': H_165_165 '5': H_168_168 '6': H_170_170 '7': H_173_173 '8': H_175_175 '9': H_178_178 '10': H_180_191 - name: weight_cls dtype: class_label: names: '0': W_38_53 '1': W_54_58 '2': W_59_62 '3': W_63_66 '4': W_67_70 '5': W_71_75 '6': W_76_79 '7': W_80_86 '8': W_87_95 '9': W_96_109 '10': W_110_186 - name: bmi_cls dtype: class_label: names: '0': BMI_12.5_20 '1': BMI_20.5_21.5 '2': BMI_22_23 '3': BMI_23.5_24.5 '4': BMI_25_26 '5': BMI_26.5_28 '6': BMI_28.5_30 '7': BMI_30.5_33.5 '8': BMI_34_38 '9': BMI_38.5_72.5 - name: pants_cls dtype: class_label: names: '0': PANTS_28_32 '1': PANTS_34_34 '2': PANTS_36_36 '3': PANTS_38_38 '4': PANTS_40_40 '5': PANTS_42_42 '6': PANTS_44_46 '7': PANTS_48_76 - name: shirt_cls dtype: class_label: names: '0': SHIRT_32_36 '1': SHIRT_40_40 '2': SHIRT_44_44 '3': SHIRT_46_46 '4': SHIRT_48_48 '5': SHIRT_50_66 - name: labels list: string splits: - name: train num_bytes: 1754024153 num_examples: 26817 - name: test num_bytes: 193214537 num_examples: 2977 download_size: 1711984856 dataset_size: 1947238690 configs: - config_name: default data_files: - split: train path: data/train-* - split: test path: data/test-* tags: - weight-estimation - height-estimation size_categories: - 10K>> ds DatasetDict({ train: Dataset({ features: ['id', 'image', 'height', 'weight', 'body_type', 'shirt_size', 'pant_size', 'body_pose_params', 'pred_global_rots', 'focal_length', 'pred_joint_coords', 'global_rot', 'shape_params', 'pred_cam_t', 'pred_keypoints_2d', 'pred_keypoints_3d', 'bbox', 'scale_params', 'bmi', 'height_cls', 'weight_cls', 'bmi_cls', 'pants_cls', 'shirt_cls', 'labels'], num_rows: 26817 }) test: Dataset({ features: ['id', 'image', 'height', 'weight', 'body_type', 'shirt_size', 'pant_size', 'body_pose_params', 'pred_global_rots', 'focal_length', 'pred_joint_coords', 'global_rot', 'shape_params', 'pred_cam_t', 'pred_keypoints_2d', 'pred_keypoints_3d', 'bbox', 'scale_params', 'bmi', 'height_cls', 'weight_cls', 'bmi_cls', 'pants_cls', 'shirt_cls', 'labels'], num_rows: 2977 }) }) ``` ### Features ![feature_importance](https://cdn-uploads.huggingface.co/production/uploads/6557b9b4deee83130ac92941/qVhzsdVL-uFd63eLgOue5.png) ![feature_correlations](https://cdn-uploads.huggingface.co/production/uploads/6557b9b4deee83130ac92941/-xrrLrF6FYUFiBLoMajUF.png) ```python >>> ds["train"].features { 'id': Value('string'), 'image': Image(mode=None, decode=True), 'height': Value('int16'), 'weight': Value('int16'), 'body_type': Value('string'), 'shirt_size': Value('string'), 'pant_size': Value('string'), 'body_pose_params': List(Value('float64')), 'pred_global_rots': List(List(List(Value('float64')))), 'focal_length': Value('float64'), 'pred_joint_coords': List(List(Value('float64'))), 'global_rot': List(Value('float64')), 'shape_params': List(Value('float64')), 'pred_cam_t': List(Value('float64')), 'pred_keypoints_2d': List(List(Value('float64'))), 'pred_keypoints_3d': List(List(Value('float64'))), 'bbox': List(Value('float64')), 'scale_params': List(Value('float64')), 'bmi': Value('float64'), 'height_cls': ClassLabel(names=['H_132_155', 'H_157_157', 'H_160_160', 'H_163_163', 'H_165_165', 'H_168_168', 'H_170_170', 'H_173_173', 'H_175_175', 'H_178_178', 'H_180_191']), 'weight_cls': ClassLabel(names=['W_38_53', 'W_54_58', 'W_59_62', 'W_63_66', 'W_67_70', 'W_71_75', 'W_76_79', 'W_80_86', 'W_87_95', 'W_96_109', 'W_110_186']), 'bmi_cls': ClassLabel(names=['BMI_12.5_20', 'BMI_20.5_21.5', 'BMI_22_23', 'BMI_23.5_24.5', 'BMI_25_26', 'BMI_26.5_28', 'BMI_28.5_30', 'BMI_30.5_33.5', 'BMI_34_38', 'BMI_38.5_72.5']), 'pants_cls': ClassLabel(names=['PANTS_28_32', 'PANTS_34_34', 'PANTS_36_36', 'PANTS_38_38', 'PANTS_40_40', 'PANTS_42_42', 'PANTS_44_46', 'PANTS_48_76']), 'shirt_cls': ClassLabel(names=['SHIRT_32_36', 'SHIRT_40_40', 'SHIRT_44_44', 'SHIRT_46_46', 'SHIRT_48_48', 'SHIRT_50_66']), 'labels': List(Value('string')) } ``` ### Sample ```python >>> ds['train'][0] id: 382327 image: height: 160 weight: 50 body_type: rectangle shirt_size: 36 pant_size: 28 body_pose_params: [[ 0.03886835 -0.00208342 -0.02778961]...[0. 0.]] shape=(133,), dtype=float64 pred_global_rots: [[[[ 1. 0. 0. ] [ 0. 1. 0. ] [ 0. 0. 1. ]] [[ 0.99640441 0.05728238 -0.06242636] [-0.06423091 0.99120814 -0.11567552] [ 0.05525135 0.1192693 0.99132341]] [[-0.0984109 -0.05719048 0.99350119] [-0.99246413 0.07888412 -0.09376723] [-0.07300889 -0.99524194 -0.06452253]]]...[[[ 0.98913765 0.07291167 0.12763445] [-0.02712706 0.94393992 -0.32900089] [-0.14446725 0.3219648 0.93566442]] [[ 0.07291168 0.03137279 0.99684489] [ 0.94393992 -0.32483506 -0.05881885] [ 0.32196486 0.94525027 -0.05329826]]]] shape=(127, 3, 3), dtype=float64 focal_length: 1014.9290161132812 pred_joint_coords: [[[ 0. -0. -0. ] [ 0. -0.92398697 -0. ] [ 0.08633898 -0.8986935 -0.00758996]]...[[ 0.0923527 -1.56066072 -0.27626431] [ 0.0634833 -1.68751919 -0.22902386]]] shape=(127, 3), dtype=float64 global_rot: [0.11973767727613449, -0.05527949333190918, -0.06437363475561142] shape_params: [[-0.76408136 0.79042131 -1.15044034]...[ 0.12328251 -0.02765466]] shape=(45,), dtype=float64 pred_cam_t: [-0.006475623697042465, 1.0043303966522217, 1.6861722469329834] pred_keypoints_2d: [[[286.81719971 143.6751709 ] [311.28182983 119.13873291] [263.91549683 114.87661743]]...[[165.12356567 245.27871704] [272.43045044 236.9727478 ]]] shape=(70, 2), dtype=float64 pred_keypoints_3d: [[[ 0.05692464 -1.51793671 -0.29545864] [ 0.09171973 -1.55984902 -0.27438799] [ 0.02577711 -1.56417561 -0.27841634]]...[[-0.12643629 -1.43218017 -0.09684809] [ 0.04102116 -1.43791533 -0.12306281]]] shape=(70, 3), dtype=float64 bbox: [54.824241638183594, 207.05712890625, 494.9328308105469, 1034.299072265625] scale_params: [[ 0.00039525 -0.0016843 -0.00043011]...[ 0.39154088 -0.13409461]] shape=(28,), dtype=float64 bmi: 19.5 height_cls: 2 weight_cls: 0 bmi_cls: 0 pants_cls: 0 shirt_cls: 0 labels: ['H_160_160', 'W_38_53', 'BMI_12.5_20', 'PANTS_28_32', 'SHIRT_32_36'] ``` ## Bins ![class_distributions](https://cdn-uploads.huggingface.co/production/uploads/6557b9b4deee83130ac92941/UuFvtlDYxPuncxZkAwwyU.png) ### Bins for height - H_132_155: 2,088 samples (7.8%) - H_157_157: 2,042 samples (7.6%) - H_160_160: 2,846 samples (10.6%) - H_163_163: 3,602 samples (13.4%) - H_165_165: 3,272 samples (12.2%) - H_168_168: 3,466 samples (12.9%) - H_170_170: 3,175 samples (11.8%) - H_173_173: 2,380 samples (8.9%) - H_175_175: 1,725 samples (6.4%) - H_178_178: 1,154 samples (4.3%) - H_180_191: 1,067 samples (4.0%) ### Bins for weight - W_38_53: 2,529 samples (9.4%) - W_54_58: 2,496 samples (9.3%) - W_59_62: 2,392 samples (8.9%) - W_63_66: 2,562 samples (9.6%) - W_67_70: 2,601 samples (9.7%) - W_71_75: 2,681 samples (10.0%) - W_76_79: 1,869 samples (7.0%) - W_80_86: 2,645 samples (9.9%) - W_87_95: 2,335 samples (8.7%) - W_96_109: 2,389 samples (8.9%) - W_110_186: 2,318 samples (8.6%) ### Bins for bmi - BMI_12.5_20: 3,043 samples (11.3%) - BMI_20.5_21.5: 2,567 samples (9.6%) - BMI_22_23: 2,743 samples (10.2%) - BMI_23.5_24.5: 2,809 samples (10.5%) - BMI_25_26: 2,666 samples (9.9%) - BMI_26.5_28: 2,716 samples (10.1%) - BMI_28.5_30: 2,243 samples (8.4%) - BMI_30.5_33.5: 2,904 samples (10.8%) - BMI_34_38: 2,476 samples (9.2%) - BMI_38.5_72.5: 2,650 samples (9.9%) ### Bins for pant_size - PANTS_28_32: 4,858 samples (18.1%) - PANTS_34_34: 3,020 samples (11.3%) - PANTS_36_36: 3,467 samples (12.9%) - PANTS_38_38: 3,431 samples (12.8%) - PANTS_40_40: 3,331 samples (12.4%) - PANTS_42_42: 2,715 samples (10.1%) - PANTS_44_46: 3,890 samples (14.5%) - PANTS_48_76: 2,105 samples (7.8%) ### Bins for shirt_size - SHIRT_32_36: 7,034 samples (26.2%) - SHIRT_40_40: 8,134 samples (30.3%) - SHIRT_44_44: 5,542 samples (20.7%) - SHIRT_46_46: 3,171 samples (11.8%) - SHIRT_48_48: 1,158 samples (4.3%) - SHIRT_50_66: 1,778 samples (6.6%) ## Analysis ### Analyses Included | Analysis | Purpose | Architecture/Training Impact | |----------|---------|------------------------------| | **Feature Structure** | Exact shapes of each feature (e.g., keypoints Nx3) | Determines if you need attention/conv vs MLP | | **Feature Statistics** | Mean, std, skewness, outliers, constant dims | Normalization strategy, dead feature pruning | | **Feature Correlations** | Inter-group correlations | Feature fusion strategy, redundancy removal | | **PCA Analysis** | Intrinsic dimensionality | Bottleneck layer sizing | | **Feature Redundancy** | Effective rank per feature group | Per-group compression potential | | **Target Correlations** | Spearman + mutual info between targets | Shared representation depth, MTL benefits | | **Class Distributions** | Imbalance ratios, entropy | Class weights, sampling strategy | | **Joint Distribution** | Multi-label co-occurrence | Combined loss behavior | | **Feature Importance** | RF importance + linear probes | Feature weighting, gating | | **Feature Ablation** | Leave-one-out accuracy drop | Critical features identification | | **Per-Feature Accuracy** | Individual group predictive power | Which features to prioritize | | **Class Separability** | LDA + t-SNE | Expected ceiling, nonlinearity needs | | **Sample Difficulty** | k-NN consistency | Curriculum learning, hard mining | | **Train/Test Comparison** | Distribution shift | Generalization expectations | | **Ordinal Structure** | Monotonicity with raw values | Ordinal regression vs classification | | **Keypoint Correlations** | Which keypoint dims predict what | Body-aware feature selection | | **Normalization Strategy** | Outliers, bounds checking | Per-feature normalization choice | | **Class Weights** | Balanced + effective weights | Loss weighting values | ### Key Questions Answered 1. **How many pose feature dimensions do you actually have?** (for MLP sizing) 2. **What's the effective dimensionality after PCA?** (bottleneck size) 3. **Which targets are correlated?** (shared representation design) 4. **How hard is each classification task?** (loss weighting) 5. **Which features matter most?** (attention/gating design) 6. **Should you use ordinal regression?** (loss function choice) 7. **How much class imbalance?** (sampling strategy) 8. **Are there redundant dimensions to prune?** (efficiency) ### Full Analysis can be found in `./analysis/analysis_results.json` ## Addendum ### Per Feature Accuracy ![per_feature_accuracy](https://cdn-uploads.huggingface.co/production/uploads/6557b9b4deee83130ac92941/02mXEZotU2JvpWNIbPabs.png) ### PCA Analysis ![pca_analysis](https://cdn-uploads.huggingface.co/production/uploads/6557b9b4deee83130ac92941/s8u8-upllRIBOAHUlAZXw.png) ### Target Correlations ![target_correlations](https://cdn-uploads.huggingface.co/production/uploads/6557b9b4deee83130ac92941/ux8LOwjun1wmStYpAJL8N.png) ### TSNE Visualization ![tsne_visualization](https://cdn-uploads.huggingface.co/production/uploads/6557b9b4deee83130ac92941/47vsYUFU7Fndw0VsbOOLk.png)