Depth from a single image, no labels needed

What it does Monodepth2 trains convolutional networks to predict depth from a single RGB image. It supports monocular video, stereo pairs, or both, and ships with pretrained KITTI models for immediate inference. The codebase covers training, finetuning, and benchmark evaluation, assuming you can supply the dataset.

The interesting bit The trick is self-supervision: it trains on image sequences without ground-truth depth labels. The README also flags a subtle evaluation quirk—stereo models are trained with an effective baseline of 0.1 units, so raw outputs need a 5.4× scaling correction to match real-world KITTI metrics.

Key highlights

• Trains without labeled depth using monocular, stereo, or combined supervision.
• Pretrained models at 640×192 and 1024×320, with and without ImageNet weights.
• Custom datasets supported by subclassing MonoDataset.
• Runs on a single GPU; authors used a Titan Xp with 6–11 GB VRAM and 8–15 hour runtimes.
• Evaluation handles multiple KITTI splits and can benchmark external disparity results.

Caveats

• Non-commercial license only.
• Single-GPU training is mandatory; no multi-GPU support.
• Replicating exact scores is uncertain due to randomization, and KITTI prep is brittle—175 GB of raw data with finicky JPEG chroma subsampling requirements.

Verdict A solid starting point for researchers and robotics tinkerers exploring self-supervised depth. Skip it if you need a commercial license or a plug-and-play product pipeline.