VINAY-UMRETHE/SigMamba-V1-Small

SigMamba: Unified Video Anomaly Detection

Weakly Supervised Video Anomaly Detection using SigLIP 2 + Mamba SSM

A unified architecture that combines SigLIP 2 (Google's SOTA vision encoder) with Mamba (Linear-complexity State Space Model) for detecting anomalies in surveillance videos. The system achieves linear O(N) scaling, enabling processing of long-form video content that was previously impractical with quadratic-cost Transformers.

---

Model Comparison

Which version should I use?

• SigMamba-V1-Large: Best for complex visual reasoning and maximum performance.
• SigMamba-V1-Small: Best for speed and lower memory usage.

Benchmark

Metric	V1 (Large)	V1 (Small)
AUC-ROC	89.82%	87.57%
Average Precision	41.05%	32.04%
Best F1-Score	41.18%	41.90%
Inference FPS	1,022	3,242
Peak VRAM	5,148 MB	3,207 MB

---

Key Features

• Linear Complexity: O(N) scaling via Mamba SSM (vs O(N²) for Transformers)
• Dual Input Modes: Accepts raw pixels or pre-extracted features

---

Architecture

The model operates in two modes:

Mode	Input	Use Case
Unified	pixel_values (B, T, 3, 384, 384)	End-to-end inference
Modular	features (B, T, 768)	Training / batch processing

---

Hyperparameters

Parameter	Value	Description
Feature Dim	768	SigLIP output dimension
Mamba d_model	768	Internal hidden dimension
Mamba Depth	8	Number of stacked layers

---

Usage

Prerequisites

pip install opencv-python
pip install transformers==4.57.3

It's recommended to use num_frames=32 due to model's training.

Loading the Model

from transformers import AutoModel, AutoProcessor
import torch

# Load the unified exported model
model = AutoModel.from_pretrained(
    "VINAY-UMRETHE/SigMamba-V1-Small",
    trust_remote_code=True
)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
model.eval()

# Load vision processor for pixel preprocessing
processor = AutoProcessor.from_pretrained(model.config.vision_model_id)

---

Inference Mode 1: Unified (Raw Pixels → Scores)

Use this when you have raw video frames. The model handles feature extraction internally.

Input Shape

pixel_values: (Batch, Time, Channels, Height, Width)
              (B, T, 3, 384, 384)

Example: Single Video

import cv2
import numpy as np

def load_video_frames(video_path, num_frames=32):
    """Sample frames uniformly from a video."""
    cap = cv2.VideoCapture(video_path)
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    indices = np.linspace(0, total_frames - 1, num_frames, dtype=int)
    
    frames = []
    for idx in indices:
        cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
        ret, frame = cap.read()
        if ret:
            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
            frames.append(frame)
    cap.release()
    return frames

# Load and preprocess
frames = load_video_frames("test_video.mp4", num_frames=32)
inputs = processor(images=frames, return_tensors="pt")
pixel_values = inputs.pixel_values.to(device)  # (32, 3, 384, 384)

# Add batch dimension: (1, 32, 3, 384, 384)
pixel_values = pixel_values.unsqueeze(0)

# Inference
with torch.no_grad():
    scores = model(pixel_values=pixel_values)
    # scores shape: (1, 32, 1)

# Get results
anomaly_scores = scores.squeeze().cpu().numpy()
max_score = anomaly_scores.max()
print(f"Max Anomaly Score: {max_score:.4f}")

---

Inference Mode 2: Modular (Pre-Extracted Features → Scores)

Use this when you've already extracted features (e.g., from batch processing).

Input Shape

features: (Batch, Time, FeatureDim)
          (B, T, 768)

Example: From Feature File

def load_features_from_txt(feature_path):
    """Load features from text file (one line per segment)."""
    with open(feature_path, 'r') as f:
        lines = f.readlines()
    features = []
    for line in lines:
        values = [float(v) for v in line.strip().split()]
        features.append(values)
    return torch.tensor(features, dtype=torch.float32)

# Load features
features = load_features_from_txt("video_features.txt")  # (T, 768)
features = features.unsqueeze(0).to(device)  # (1, T, 768)

# Inference
with torch.no_grad():
    scores = model(features=features)
    # scores shape: (1, T, 1)

print(f"Anomaly Scores: {scores.squeeze().cpu().numpy()}")

---

Batch Processing Multiple Videos

Process multiple videos in a single forward pass for efficiency.

# Load multiple videos
video_paths = ["video1.mp4", "video2.mp4", "video3.mp4"]
batch_frames = []

for path in video_paths:
    frames = load_video_frames(path, num_frames=32)
    inputs = processor(images=frames, return_tensors="pt")
    batch_frames.append(inputs.pixel_values)

# Stack into batch: (3, 32, 3, 384, 384)
pixel_values = torch.stack(batch_frames).to(device)

# Single forward pass for all videos
with torch.no_grad():
    scores = model(pixel_values=pixel_values)
    # scores shape: (3, 32, 1)

# Per-video max scores
for i, path in enumerate(video_paths):
    max_score = scores[i].max().item()
    print(f"{path}: {max_score:.4f}")

---

Single Frame Analysis

For quick spot-checks on individual frames.

from PIL import Image

# Load single image
image = Image.open("suspicious_frame.jpg")
inputs = processor(images=image, return_tensors="pt")
pixel_values = inputs.pixel_values.to(device)  # (1, 3, 384, 384)

# Reshape: (1, 1, 3, 384, 384) - batch=1, time=1
pixel_values = pixel_values.unsqueeze(0)

with torch.no_grad():
    score = model(pixel_values=pixel_values)
    print(f"Frame Anomaly Score: {score.item():.4f}")

---

Extract Features Only (No Classification)

Access the Mamba encoder output directly for custom downstream tasks.

# Load frames
frames = load_video_frames("video.mp4", num_frames=32)
inputs = processor(images=frames, return_tensors="pt")
pixel_values = inputs.pixel_values.unsqueeze(0).to(device)

# Access internal components
with torch.no_grad():
    # Step 1: Extract vision features
    b, t, c, h, w = pixel_values.shape
    flat_pixels = pixel_values.view(b * t, c, h, w)
    vision_features = model.vision_model.get_image_features(pixel_values=flat_pixels)
    vision_features = vision_features / vision_features.norm(dim=-1, keepdim=True)
    vision_features = vision_features.view(b, t, -1)  # (1, 32, 768)
    
    # Step 2: Get Mamba-encoded features
    mamba_features = model.mamba_encoder(vision_features)  # (1, 32, 768)
    
    print(f"Vision Features: {vision_features.shape}")
    print(f"Mamba Features: {mamba_features.shape}")

---

Threshold-Based Detection

Apply a threshold to convert scores into binary predictions.

def detect_anomalies(video_path, threshold=0.5):
    """Returns list of anomalous segment indices."""
    frames = load_video_frames(video_path, num_frames=32)
    inputs = processor(images=frames, return_tensors="pt")
    pixel_values = inputs.pixel_values.unsqueeze(0).to(device)
    
    with torch.no_grad():
        scores = model(pixel_values=pixel_values)
        scores = scores.squeeze().cpu().numpy()
    
    anomalous_segments = np.where(scores > threshold)[0]
    
    return {
        "scores": scores,
        "max_score": scores.max(),
        "is_anomalous": scores.max() > threshold,
        "anomalous_segments": anomalous_segments.tolist()
    }

# Usage
result = detect_anomalies("test.mp4", threshold=0.5)
print(f"Anomalous: {result['is_anomalous']}")
print(f"Segments: {result['anomalous_segments']}")

Output Reference

Method	Input	Output Shape	Description
model(pixel_values=...)	(B, T, 3, 384, 384)	(B, T, 1)	End-to-end inference
model(features=...)	(B, T, 768)	(B, T, 1)	Feature-based inference
model.mamba_encoder(...)	(B, T, 768)	(B, T, 768)	Encoded temporal features
model.vision_model.get_image_features(...)	(N, 3, 384, 384)	(N, 768)	Raw vision embeddings

---

License

This model is licensed under the MIT License.