Simuletic commited on
Commit
754de50
·
verified ·
1 Parent(s): 011435b

Simuletic Fall Detector v1.0 (Nano recommended, Small, ONNX)

Browse files
.gitattributes CHANGED
@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ assets/hero_synthetic_grid.jpg filter=lfs diff=lfs merge=lfs -text
37
+ assets/validation.gif filter=lfs diff=lfs merge=lfs -text
38
+ assets/validation.mp4 filter=lfs diff=lfs merge=lfs -text
39
+ assets/validation.webp filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,181 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ library_name: transformers
4
+ pipeline_tag: object-detection
5
+ tags:
6
+ - object-detection
7
+ - d-fine
8
+ - cctv
9
+ - fall-detection
10
+ - person-on-floor
11
+ - synthetic-data
12
+ - onnx
13
+ datasets:
14
+ - Simuletic/CCTV_Incident_Dataset_Fall_Lying_Down_Detection
15
+ base_model:
16
+ - ustc-community/dfine-small-coco
17
+ - ustc-community/dfine-nano-coco
18
+ ---
19
+
20
+ # Simuletic Fall Detector
21
+
22
+ **Detects a person lying on the floor after a fall, with a 2-second stays-down alarm.**
23
+
24
+ The model finds a **person lying on the ground or floor** (fallen, collapsed) in CCTV-style frames and tells them apart from
25
+ people who are standing, sitting, bending, kneeling or crouching. **Trained only on synthetic images.** No real
26
+ people or real footage were used for training. Tested on real, staged footage from two homes the model has never seen.
27
+
28
+ ![The release model on real test footage it never saw: a fall raises the 2-second alarm, a pick-up does not, and a long, deep bend correctly does not](assets/validation.webp)
29
+
30
+ *The release model on **real** test footage it never saw in training (GMDCSA-24, MIT; CAUCAFall, CC BY 4.0). Model: Nano (recommended). Clips are chosen by a fixed rule (the median-scoring fall in each test set, an everyday pick-up, and the hardest look-alike in the tests: a deep bend held for more than 2 s, which Nano correctly ignores and Small does not). Full video: [validation.mp4](assets/validation.mp4).*
31
+
32
+ ![Synthetic training data: fallen people (red) and look-alikes such as bending, kneeling and sitting (blue), day and night infrared](assets/hero_synthetic_grid.jpg)
33
+
34
+ *Synthetic training data (Simuletic). Red = `fallen_person`, blue = `person`.*
35
+
36
+ | | Small | **Nano (recommended)** |
37
+ |---|---|---|
38
+ | Architecture | D-FINE S | D-FINE N |
39
+ | Parameters | 10.2 M | 3.7 M |
40
+ | Input | 640 × 640 | 640 × 640 |
41
+ | Speed, Apple M4 Pro, batch 1, PyTorch incl. pre/post-processing | GPU 22 ms, CPU 132 ms | GPU 16 ms, CPU 94 ms |
42
+ | Files | `small/model.safetensors`, `small/model.onnx` | `nano/model.safetensors`, `nano/model.onnx` |
43
+ | Classes | 0 `person` (upright), 1 `fallen_person` (on the floor, not holding themselves up) | same |
44
+
45
+ **Licence: CC BY-NC 4.0, for non-commercial research and evaluation only.** For commercial use, see
46
+ [Commercial use and the Pro model](#commercial-use-and-the-pro-model).
47
+
48
+ ## Results on real footage
49
+
50
+ Every number is the **mean of 2 training seeds** (range in brackets), never the best seed. The model has seen
51
+ no frame from either test set.
52
+
53
+ **Test 1: CAUCAFall**, subjects 6 to 10. One home, single frames; the person is scored once they are on the floor.
54
+
55
+ | | Small | Nano |
56
+ |---|---|---|
57
+ | Frame AP | 0.975 (0.970–0.979) | 0.995 (0.994–0.997) |
58
+ | Lying frames caught, 5% of other frames alarming | 99.8% (99.7–100.0%) | 99.3% (99.0–99.7%) |
59
+ | Lying frames caught, 1% of other frames alarming | 79.8% (74.2–85.5%) | 95.6% (94.8–96.1%) |
60
+ | End-on lying (head or feet towards the camera), 5% | 100.0% | 100.0% |
61
+
62
+ **Test 2: GMDCSA-24.** A different home and different people, scored per clip.
63
+
64
+ | | Small | Nano |
65
+ |---|---|---|
66
+ | Clip AP | 0.981 (0.977–0.985) | 0.987 (0.986–0.989) |
67
+ | Falls caught, 10% of everyday clips alarming | 96.8% (96.2–97.5%) | 94.5% (93.7–96.2%) |
68
+
69
+ **"Stays down for 2 seconds" alarm** (Test 1). The alarm fires only when every frame in the last 2 s agrees.
70
+
71
+ | | Small | Nano |
72
+ |---|---|---|
73
+ | Falls caught with 0 false-alarm events | 60.4% (37.5–83.3%) | 100.0% |
74
+ | Falls caught with 1 false-alarm event | 100.0% | 100.0% |
75
+
76
+ ### Why synthetic
77
+
78
+ Trained on real footage from one home, the same detector (D-FINE Nano) is perfect on more footage from that home, but in
79
+ the second home it catches only **24%** of falls. Trained only on synthetic images, it catches **94.5% (93.7–96.2%)** (Small: **96.8% (96.2–97.5%)**). Real footage from one place teaches that place; synthetic data can cover rooms, cameras, lighting,
80
+ people and look-alikes that no single recording has.
81
+
82
+ ## Use
83
+
84
+ ```python
85
+ from PIL import Image
86
+ from transformers import AutoImageProcessor, AutoModelForObjectDetection
87
+
88
+ repo = "Simuletic/fall-detector"
89
+ processor = AutoImageProcessor.from_pretrained(repo, subfolder="nano")
90
+ model = AutoModelForObjectDetection.from_pretrained(repo, subfolder="nano").eval()
91
+
92
+ image = Image.open("frame.jpg").convert("RGB")
93
+ outputs = model(**processor(images=image, return_tensors="pt"))
94
+ result = processor.post_process_object_detection(outputs, target_sizes=[(image.height, image.width)], threshold=0.5)[0]
95
+ for score, label, box in zip(result["scores"], result["labels"], result["boxes"]):
96
+ print(model.config.id2label[int(label)], round(float(score), 2), box.tolist())
97
+ ```
98
+
99
+ The helper package has the "stays down" alarm, video, ONNX and the evaluation scripts:
100
+ [github.com/Simuletic/fall-detector](https://github.com/Simuletic/fall-detector).
101
+ Try it in the [Space](https://huggingface.co/spaces/Simuletic/fall-detector) or the
102
+ [Kaggle notebook](https://www.kaggle.com/code/simuletic/fall-detection-on-real-footage-trained-only-on-synthetic-data).
103
+
104
+ **ONNX:** input `pixel_values` [1, 3, 640, 640], RGB, resized to 640 × 640 and scaled to 0–1 (no mean/std).
105
+ The outputs are `logits` [1, 300, 2] (apply a sigmoid per class) and `pred_boxes` [1, 300, 4] (cx, cy, w, h, normalised).
106
+
107
+ **Thresholds:** 0.5 is a starting point. Choose your own on footage from your cameras.
108
+
109
+ ## How the numbers were measured
110
+
111
+ - **Test 1, CAUCAFall** (CC BY 4.0), subjects 6 to 10:
112
+ - Positives are frames from 1 s after the fall onset, with the person on the floor.
113
+ - Negatives are walking, hopping, sitting down, kneeling and picking up an object, plus the fall clips before the onset.
114
+ - The second either side of the onset (mid-fall) is left out.
115
+ - 1,923 frames, 615 lying.
116
+ - **Test 2, GMDCSA-24** (MIT), a different home with 4 people:
117
+ - 2 frames per second; a clip's score is its second-highest frame score.
118
+ - Lying in bed or on a sofa is out of scope, so those 41 clips are left out.
119
+ - **Stays down:** persistent score = lowest frame score over the last 2 s. Test 1 has 24 fall events and 34 negative
120
+ events, so one event is about 4 points.
121
+ - **Seeds:** 2 for Small and 3 for Nano, same recipe. We report the mean and the range, not the best seed.
122
+ - Scripts: [eval/](https://github.com/Simuletic/fall-detector/tree/main/eval).
123
+
124
+ ## Training
125
+
126
+ - **Data: synthetic only, about 1,580 images.**
127
+ - Simuletic's [CCTV Fall Detection dataset](https://simuletic.com/datasets?utm_source=huggingface&utm_medium=model_card&utm_campaign=fall-detector) (1,000 images).
128
+ - A 438-image top-up: people lying end-on to the camera, partly hidden by furniture, curled and slumped; chair, sit-down, bend and kneel negatives.
129
+ - A 138-image round of look-alikes: deep bends to pick something up, squatting, kneeling on all fours, reaching under furniture, crouching to tie shoes. About half are night infrared, in homes, care homes, shops, car parks and corridors.
130
+ - Fixed CCTV-style cameras 2.5–7 m high.
131
+ - **Recipe:**
132
+ - Starts from the D-FINE COCO checkpoints.
133
+ - 60 epochs at 640 px, batch 8, AdamW.
134
+ - Learning rate 2.5e-4, 10× lower for the pretrained feature extractor; one-cycle schedule.
135
+ - Trained on an Apple M4 Pro.
136
+
137
+ ## Limitations
138
+
139
+ - **Small, staged, indoor tests.** CAUCAFall has 5 test people in one home; GMDCSA-24 has 4 people in one home. Both are
140
+ staged. We have **not measured** public spaces, crowds, long distances or real night-time infrared CCTV.
141
+ - **Out of scope:** lying in bed or on a sofa, and the falling motion itself (the model scores single frames).
142
+ - **Hardest look-alike:** a long, deep bend to pick something up off the floor. On Test 1, one person bends double for more than 4 s with the head towards the camera. Nano keeps this below the alarm on all 3 seeds (single frames peak around 0.8, never for 2 s). **Small still alarms on it** on both seeds, so prefer Nano where look-alikes matter. Long, deep bends in other homes and camera angles are untested.
143
+ - **People partly out of frame or heavily occluded** were only covered up to about half of the body.
144
+ - **Not a medical device or a safety system.** Do not rely on it to protect anyone.
145
+ - **Fairness:** the training images vary age, gender, build and skin tone by design, but the real test sets are too small
146
+ to measure differences between groups.
147
+
148
+ ## Commercial use and the Pro model
149
+
150
+ These weights are licensed for **non-commercial research and evaluation** only (CC BY-NC 4.0). Products, pilots, paid
151
+ services and any other commercial use need a licence.
152
+
153
+ **Simuletic Pro** is trained on request for your product:
154
+
155
+ - 10k+ training images (more night infrared, camera heights and angles, look-alikes);
156
+ - edge builds;
157
+ - a commercial licence and support;
158
+ - optional fine-tuning on your own camera views.
159
+
160
+ Contact: [simuletic.com/contact](https://simuletic.com/contact?utm_source=huggingface&utm_medium=model_card&utm_campaign=fall-detector).
161
+
162
+ ## Credits and licences
163
+
164
+ - **Model:** D-FINE (Apache-2.0), via Hugging Face `transformers`, fine-tuned from `ustc-community/dfine-small-coco` and
165
+ `ustc-community/dfine-nano-coco` (COCO pretraining).
166
+ - **Training images:** fully synthetic, made by Simuletic.
167
+ - **Test sets, used for evaluation only (no frames used for training):**
168
+ - CAUCAFall: Eraso Guerrero et al., *Data in Brief* 2022, [doi:10.17632/7w7fccy7ky](https://data.mendeley.com/datasets/7w7fccy7ky), CC BY 4.0.
169
+ - GMDCSA-24: Alam et al., [GitHub](https://github.com/ekramalam/GMDCSA24-A-Dataset-for-Human-Fall-Detection-in-Videos), MIT.
170
+ - **Weights:** © Simuletic, CC BY-NC 4.0.
171
+
172
+ ## Citation
173
+
174
+ ```bibtex
175
+ @misc{simuletic_fall_detector_2026,
176
+ title = {Simuletic Fall Detector: a person-on-the-floor detector trained only on synthetic CCTV images},
177
+ author = {{Simuletic}},
178
+ year = {2026},
179
+ url = {https://huggingface.co/Simuletic/fall-detector}
180
+ }
181
+ ```
assets/before_after.jpg ADDED
assets/hero_synthetic_grid.jpg ADDED

Git LFS Details

  • SHA256: a95224f640a5560e05bab3c9b1d9d12b0280d22b4214c626bd04ea87220f3569
  • Pointer size: 131 Bytes
  • Size of remote file: 258 kB
assets/how_it_works.png ADDED
assets/validation.gif ADDED

Git LFS Details

  • SHA256: 47c2d8bb00428cc0e5bcec934874c324b0c64a6135abb47366562e91dd241b99
  • Pointer size: 132 Bytes
  • Size of remote file: 3.8 MB
assets/validation.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:11b2edf6ae294ef5a0851c80f5a50eaee2e4b307cbe9bc2237717c8b040a5fb1
3
+ size 1649557
assets/validation.webp ADDED

Git LFS Details

  • SHA256: 0321f85df189be94b7331d36b6ceec54c714151be38fc747f6c06da377ed4f93
  • Pointer size: 132 Bytes
  • Size of remote file: 1.56 MB
nano/config.json ADDED
@@ -0,0 +1,217 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "activation_dropout": 0.0,
3
+ "activation_function": "silu",
4
+ "anchor_image_size": null,
5
+ "architectures": [
6
+ "DFineForObjectDetection"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "auxiliary_loss": true,
10
+ "backbone": null,
11
+ "backbone_config": {
12
+ "_name_or_path": "",
13
+ "architectures": null,
14
+ "chunk_size_feed_forward": 0,
15
+ "depths": [
16
+ 3,
17
+ 4,
18
+ 6,
19
+ 3
20
+ ],
21
+ "downsample_in_bottleneck": false,
22
+ "downsample_in_first_stage": false,
23
+ "dtype": "float32",
24
+ "embedding_size": 32,
25
+ "hidden_act": "relu",
26
+ "hidden_sizes": [
27
+ 128,
28
+ 256,
29
+ 512,
30
+ 1024
31
+ ],
32
+ "id2label": {
33
+ "0": "LABEL_0",
34
+ "1": "LABEL_1"
35
+ },
36
+ "initializer_range": 0.02,
37
+ "is_encoder_decoder": false,
38
+ "label2id": {
39
+ "LABEL_0": 0,
40
+ "LABEL_1": 1
41
+ },
42
+ "layer_type": "basic",
43
+ "model_type": "hgnet_v2",
44
+ "num_channels": 3,
45
+ "out_features": [
46
+ "stage3",
47
+ "stage4"
48
+ ],
49
+ "out_indices": [
50
+ 3,
51
+ 4
52
+ ],
53
+ "output_attentions": false,
54
+ "output_hidden_states": false,
55
+ "problem_type": null,
56
+ "return_dict": true,
57
+ "stage_downsample": [
58
+ false,
59
+ true,
60
+ true,
61
+ true
62
+ ],
63
+ "stage_downsample_strides": [
64
+ 2,
65
+ 2,
66
+ 2,
67
+ 2
68
+ ],
69
+ "stage_in_channels": [
70
+ 16,
71
+ 64,
72
+ 256,
73
+ 512
74
+ ],
75
+ "stage_kernel_size": [
76
+ 3,
77
+ 3,
78
+ 5,
79
+ 5
80
+ ],
81
+ "stage_light_block": [
82
+ false,
83
+ false,
84
+ true,
85
+ true
86
+ ],
87
+ "stage_mid_channels": [
88
+ 16,
89
+ 32,
90
+ 64,
91
+ 128
92
+ ],
93
+ "stage_names": [
94
+ "stem",
95
+ "stage1",
96
+ "stage2",
97
+ "stage3",
98
+ "stage4"
99
+ ],
100
+ "stage_num_blocks": [
101
+ 1,
102
+ 1,
103
+ 2,
104
+ 1
105
+ ],
106
+ "stage_numb_of_layers": [
107
+ 3,
108
+ 3,
109
+ 3,
110
+ 3
111
+ ],
112
+ "stage_out_channels": [
113
+ 64,
114
+ 256,
115
+ 512,
116
+ 1024
117
+ ],
118
+ "stem_channels": [
119
+ 3,
120
+ 16,
121
+ 16
122
+ ],
123
+ "stem_strides": [
124
+ 2,
125
+ 1,
126
+ 1,
127
+ 2,
128
+ 1
129
+ ],
130
+ "use_learnable_affine_block": true
131
+ },
132
+ "batch_norm_eps": 1e-05,
133
+ "box_noise_scale": 1.0,
134
+ "d_model": 128,
135
+ "decoder_activation_function": "relu",
136
+ "decoder_attention_heads": 8,
137
+ "decoder_ffn_dim": 512,
138
+ "decoder_in_channels": [
139
+ 128,
140
+ 128
141
+ ],
142
+ "decoder_layers": 3,
143
+ "decoder_method": "default",
144
+ "decoder_n_points": [
145
+ 6,
146
+ 6
147
+ ],
148
+ "decoder_offset_scale": 0.5,
149
+ "depth_mult": 0.5,
150
+ "dropout": 0.0,
151
+ "dtype": "float32",
152
+ "encode_proj_layers": [
153
+ 1
154
+ ],
155
+ "encoder_activation_function": "gelu",
156
+ "encoder_attention_heads": 8,
157
+ "encoder_ffn_dim": 512,
158
+ "encoder_hidden_dim": 128,
159
+ "encoder_in_channels": [
160
+ 512,
161
+ 1024
162
+ ],
163
+ "encoder_layers": 1,
164
+ "eos_coefficient": 0.0001,
165
+ "eval_idx": -1,
166
+ "eval_size": null,
167
+ "feat_strides": [
168
+ 16,
169
+ 32
170
+ ],
171
+ "focal_loss_alpha": 0.75,
172
+ "focal_loss_gamma": 2.0,
173
+ "freeze_backbone_batch_norms": true,
174
+ "head_dim": 16,
175
+ "hidden_expansion": 0.34,
176
+ "id2label": {
177
+ "0": "person",
178
+ "1": "fallen_person"
179
+ },
180
+ "initializer_bias_prior_prob": null,
181
+ "initializer_range": 0.01,
182
+ "is_encoder_decoder": true,
183
+ "label2id": {
184
+ "fallen_person": 1,
185
+ "person": 0
186
+ },
187
+ "label_noise_ratio": 0.5,
188
+ "layer_norm_eps": 1e-05,
189
+ "layer_scale": 1,
190
+ "learn_initial_query": false,
191
+ "lqe_hidden_dim": 64,
192
+ "lqe_layers": 2,
193
+ "matcher_alpha": 0.25,
194
+ "matcher_bbox_cost": 5.0,
195
+ "matcher_class_cost": 2.0,
196
+ "matcher_gamma": 2.0,
197
+ "matcher_giou_cost": 2.0,
198
+ "max_num_bins": 32,
199
+ "model_type": "d_fine",
200
+ "normalize_before": false,
201
+ "num_denoising": 100,
202
+ "num_feature_levels": 2,
203
+ "num_queries": 300,
204
+ "positional_encoding_temperature": 10000,
205
+ "reg_scale": 4.0,
206
+ "tie_word_embeddings": true,
207
+ "top_prob_values": 4,
208
+ "transformers_version": "5.20.2",
209
+ "up": 0.5,
210
+ "use_focal_loss": true,
211
+ "weight_loss_bbox": 5.0,
212
+ "weight_loss_ddf": 1.5,
213
+ "weight_loss_fgl": 0.15,
214
+ "weight_loss_giou": 2.0,
215
+ "weight_loss_vfl": 1.0,
216
+ "with_box_refine": true
217
+ }
nano/model.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c36f557c7351148218d7dc7e91aa6fd522ef66d94d20089d55a9e1bfdb45e49
3
+ size 15351838
nano/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:af6ee951ae0aedc518201817544e7475d1c1404e225169ad6a8970fd4054c83f
3
+ size 15078052
nano/preprocessor_config.json ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_convert_annotations": true,
3
+ "do_normalize": false,
4
+ "do_pad": false,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "format": "coco_detection",
8
+ "image_mean": [
9
+ 0.485,
10
+ 0.456,
11
+ 0.406
12
+ ],
13
+ "image_processor_type": "RTDetrImageProcessor",
14
+ "image_std": [
15
+ 0.229,
16
+ 0.224,
17
+ 0.225
18
+ ],
19
+ "resample": 2,
20
+ "rescale_factor": 0.00392156862745098,
21
+ "size": {
22
+ "height": 640,
23
+ "width": 640
24
+ }
25
+ }
small/config.json ADDED
@@ -0,0 +1,223 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "activation_dropout": 0.0,
3
+ "activation_function": "silu",
4
+ "anchor_image_size": null,
5
+ "architectures": [
6
+ "DFineForObjectDetection"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "auxiliary_loss": true,
10
+ "backbone": null,
11
+ "backbone_config": {
12
+ "_name_or_path": "",
13
+ "architectures": null,
14
+ "chunk_size_feed_forward": 0,
15
+ "depths": [
16
+ 3,
17
+ 4,
18
+ 6,
19
+ 3
20
+ ],
21
+ "downsample_in_bottleneck": false,
22
+ "downsample_in_first_stage": false,
23
+ "dtype": "float32",
24
+ "embedding_size": 32,
25
+ "hidden_act": "relu",
26
+ "hidden_sizes": [
27
+ 128,
28
+ 256,
29
+ 512,
30
+ 1024
31
+ ],
32
+ "id2label": {
33
+ "0": "LABEL_0",
34
+ "1": "LABEL_1"
35
+ },
36
+ "initializer_range": 0.02,
37
+ "is_encoder_decoder": false,
38
+ "label2id": {
39
+ "LABEL_0": 0,
40
+ "LABEL_1": 1
41
+ },
42
+ "layer_type": "basic",
43
+ "model_type": "hgnet_v2",
44
+ "num_channels": 3,
45
+ "out_features": [
46
+ "stage2",
47
+ "stage3",
48
+ "stage4"
49
+ ],
50
+ "out_indices": [
51
+ 2,
52
+ 3,
53
+ 4
54
+ ],
55
+ "output_attentions": false,
56
+ "output_hidden_states": false,
57
+ "problem_type": null,
58
+ "return_dict": true,
59
+ "stage_downsample": [
60
+ false,
61
+ true,
62
+ true,
63
+ true
64
+ ],
65
+ "stage_downsample_strides": [
66
+ 2,
67
+ 2,
68
+ 2,
69
+ 2
70
+ ],
71
+ "stage_in_channels": [
72
+ 16,
73
+ 64,
74
+ 256,
75
+ 512
76
+ ],
77
+ "stage_kernel_size": [
78
+ 3,
79
+ 3,
80
+ 5,
81
+ 5
82
+ ],
83
+ "stage_light_block": [
84
+ false,
85
+ false,
86
+ true,
87
+ true
88
+ ],
89
+ "stage_mid_channels": [
90
+ 16,
91
+ 32,
92
+ 64,
93
+ 128
94
+ ],
95
+ "stage_names": [
96
+ "stem",
97
+ "stage1",
98
+ "stage2",
99
+ "stage3",
100
+ "stage4"
101
+ ],
102
+ "stage_num_blocks": [
103
+ 1,
104
+ 1,
105
+ 2,
106
+ 1
107
+ ],
108
+ "stage_numb_of_layers": [
109
+ 3,
110
+ 3,
111
+ 3,
112
+ 3
113
+ ],
114
+ "stage_out_channels": [
115
+ 64,
116
+ 256,
117
+ 512,
118
+ 1024
119
+ ],
120
+ "stem_channels": [
121
+ 3,
122
+ 16,
123
+ 16
124
+ ],
125
+ "stem_strides": [
126
+ 2,
127
+ 1,
128
+ 1,
129
+ 2,
130
+ 1
131
+ ],
132
+ "use_learnable_affine_block": true
133
+ },
134
+ "batch_norm_eps": 1e-05,
135
+ "box_noise_scale": 1.0,
136
+ "d_model": 256,
137
+ "decoder_activation_function": "relu",
138
+ "decoder_attention_heads": 8,
139
+ "decoder_ffn_dim": 1024,
140
+ "decoder_in_channels": [
141
+ 256,
142
+ 256,
143
+ 256
144
+ ],
145
+ "decoder_layers": 3,
146
+ "decoder_method": "default",
147
+ "decoder_n_points": [
148
+ 3,
149
+ 6,
150
+ 3
151
+ ],
152
+ "decoder_offset_scale": 0.5,
153
+ "depth_mult": 0.34,
154
+ "dropout": 0.0,
155
+ "dtype": "float32",
156
+ "encode_proj_layers": [
157
+ 2
158
+ ],
159
+ "encoder_activation_function": "gelu",
160
+ "encoder_attention_heads": 8,
161
+ "encoder_ffn_dim": 1024,
162
+ "encoder_hidden_dim": 256,
163
+ "encoder_in_channels": [
164
+ 256,
165
+ 512,
166
+ 1024
167
+ ],
168
+ "encoder_layers": 1,
169
+ "eos_coefficient": 0.0001,
170
+ "eval_idx": -1,
171
+ "eval_size": null,
172
+ "feat_strides": [
173
+ 8,
174
+ 16,
175
+ 32
176
+ ],
177
+ "focal_loss_alpha": 0.75,
178
+ "focal_loss_gamma": 2.0,
179
+ "freeze_backbone_batch_norms": true,
180
+ "head_dim": 32,
181
+ "hidden_expansion": 0.5,
182
+ "id2label": {
183
+ "0": "person",
184
+ "1": "fallen_person"
185
+ },
186
+ "initializer_bias_prior_prob": null,
187
+ "initializer_range": 0.01,
188
+ "is_encoder_decoder": true,
189
+ "label2id": {
190
+ "fallen_person": 1,
191
+ "person": 0
192
+ },
193
+ "label_noise_ratio": 0.5,
194
+ "layer_norm_eps": 1e-05,
195
+ "layer_scale": 1,
196
+ "learn_initial_query": false,
197
+ "lqe_hidden_dim": 64,
198
+ "lqe_layers": 2,
199
+ "matcher_alpha": 0.25,
200
+ "matcher_bbox_cost": 5.0,
201
+ "matcher_class_cost": 2.0,
202
+ "matcher_gamma": 2.0,
203
+ "matcher_giou_cost": 2.0,
204
+ "max_num_bins": 32,
205
+ "model_type": "d_fine",
206
+ "normalize_before": false,
207
+ "num_denoising": 100,
208
+ "num_feature_levels": 3,
209
+ "num_queries": 300,
210
+ "positional_encoding_temperature": 10000,
211
+ "reg_scale": 4.0,
212
+ "tie_word_embeddings": true,
213
+ "top_prob_values": 4,
214
+ "transformers_version": "5.20.2",
215
+ "up": 0.5,
216
+ "use_focal_loss": true,
217
+ "weight_loss_bbox": 5.0,
218
+ "weight_loss_ddf": 1.5,
219
+ "weight_loss_fgl": 0.15,
220
+ "weight_loss_giou": 2.0,
221
+ "weight_loss_vfl": 1.0,
222
+ "with_box_refine": true
223
+ }
small/model.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ba331b4388d13c665260e484922df467ed479cc9aa277e31d57134fc4952b16e
3
+ size 41347921
small/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ed619092e75b6c1870c8e8c2fc890270891cb9f8e5ab9ab16f9b8768bc2aa6c6
3
+ size 41097004
small/preprocessor_config.json ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_convert_annotations": true,
3
+ "do_normalize": false,
4
+ "do_pad": false,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "format": "coco_detection",
8
+ "image_mean": [
9
+ 0.485,
10
+ 0.456,
11
+ 0.406
12
+ ],
13
+ "image_processor_type": "RTDetrImageProcessor",
14
+ "image_std": [
15
+ 0.229,
16
+ 0.224,
17
+ 0.225
18
+ ],
19
+ "resample": 2,
20
+ "rescale_factor": 0.00392156862745098,
21
+ "size": {
22
+ "height": 640,
23
+ "width": 640
24
+ }
25
+ }