Scene Understanding

Module: Computer Vision | Difficulty: Intermediate

Scene Classification

Places365

365 scene categories with CNN features.

Scene Graph Generation

Visual Relationship Detection

Scene Attribute Prediction

import torch
import torch.nn as nn

class SceneGraphHead(nn.Module):
    def __init__(self, obj_classes, rel_classes, feat_dim=256):
        super().__init__()
        self.obj_embed = nn.Embedding(obj_classes, feat_dim)
        self.rel_predictor = nn.Sequential(
            nn.Linear(feat_dim * 3, 512),
            nn.ReLU(True),
            nn.Linear(512, rel_classes)
        )
    
    def forward(self, subj_feat, obj_feat, region_feat):
        s = self.obj_embed(subj_feat)
        o = self.obj_embed(obj_feat)
        combined = torch.cat([s, o, region_feat], dim=1)
        return self.rel_predictor(combined)

Key Takeaways

• Scene understanding goes beyond object detection
• Scene graphs represent relationships between objects
• Holistic analysis combines classification, detection, and description