Curated References & Further Reading

"They cite the textbook ten thousand times and read it maybe forty. The afternoon I finally read it, my bug was waiting on page 312, in a footnote, with the sign convention spelled out."

A Pose Estimator Who Finally Opened the Book

A bad reading list buries you in fifty links you will never open; a good one tells you the single source to reach for when a specific function misbehaves. This section is the second kind. It gives you the two books that anchor the field, one canonical reference per chapter of Part II, the lecture series worth your evenings, and the documentation trails behind the tools of Section 17.1 and Section 17.2. It closes with the part of the skill no book teaches: how to stay current without drowning. Treat it as a routing table, not a syllabus.

1. The Two Books to Own Beginner

Two books carry most of Part II. Hartley and Zisserman, Multiple View Geometry in Computer Vision (2nd ed., 2004) is the definitive treatment of the projective geometry behind Chapter 12 through Chapter 14: the fundamental and essential matrices, the camera projection model, triangulation, and bundle adjustment, all with the sign conventions and degeneracy analysis that the function docstrings of Section 17.1 assume but never explain. It is dense, and you read it the way you read a reference: when recoverPose returns a rotation you did not expect, the answer is in its pages. Szeliski, Computer Vision: Algorithms and Applications (2nd ed., 2022) is the opposite personality: broad, current, free online, and readable cover to cover, with chapters that map almost one-to-one onto this book's Part II and full citation trails into the literature.

Key Insight: The Old Books Debug the New Methods

It is tempting to assume that a 2004 textbook is obsolete in a field moving as fast as vision. The opposite is true for geometry. Learned methods such as the feed-forward reconstructors of Section 17.2 still produce camera matrices, essential matrices, and reprojection errors, the same objects Hartley and Zisserman defined, and they fail in the same geometric ways: scale ambiguity, planar degeneracy, and gauge freedom. Gauge freedom means a reconstruction is fixed only up to a global similarity transform: the whole scene can be rotated, translated, and rescaled without changing any reprojection, as Section 14.3 spells out. When a neural pose estimator behaves strangely, the explanation is almost never in the network architecture and almost always in the projective geometry the network is approximating. The classical books are not history; they are the specification that the learned methods implement, and the manual for diagnosing them when they break. The phrase worth keeping: the network forgets, the geometry does not. The illustration below casts that idea as the old green book patiently debugging a modern neural reconstructor.

2. One Paper Per Chapter Intermediate

If you read one primary source for each chapter of Part II, read these. They are the papers the algorithms are named after, and going to the source resolves the ambiguities that secondary explanations introduce. Table 17.4.1 routes each chapter to its canonical paper; the chapter bibliographies hold the full hyperlinked entries.

A practical reading habit makes these sources far more useful, and it can be partly automated. Code 17.4.1 builds a tiny personal citation index: given a list of arXiv identifiers, it fetches each title and abstract so you can skim a reading queue without opening eight browser tabs.

# Personal arXiv reading queue: given a list of paper IDs, query the
# public arXiv API and print each title with a short abstract snippet,
# so a reading list can be skimmed without opening a browser tab per paper.
import urllib.request
import xml.etree.ElementTree as ET

def arxiv_titles(ids):
    """Fetch title + abstract for a list of arXiv IDs via the public API."""
    query = ",".join(ids)
    url = f"http://export.arxiv.org/api/query?id_list={query}&max_results={len(ids)}"
    with urllib.request.urlopen(url, timeout=20) as resp:
        root = ET.fromstring(resp.read())
    ns = {"a": "http://www.w3.org/2005/Atom"}
    for entry in root.findall("a:entry", ns):
        title = entry.find("a:title", ns).text.strip().replace("\n", " ")
        summary = entry.find("a:summary", ns).text.strip()[:160]
        print(f"- {title}\n    {summary}...\n")

# A starter reading queue: GLOMAP, SuperPoint, RAFT, DUSt3R, VGGT.
arxiv_titles(["2407.20219", "1712.07629", "2003.12039", "2312.14132", "2503.11651"])

- Global Structure-from-Motion Revisited
    We revisit the problem of estimating camera poses and 3D structure...

- SuperPoint: Self-Supervised Interest Point Detection and Description
    This paper presents a self-supervised framework for training interest...

- RAFT: Recurrent All-Pairs Field Transforms for Optical Flow
    We introduce Recurrent All-Pairs Field Transforms (RAFT), a new deep...

3. Lectures, Docs & a Staying-Current Workflow Intermediate

Books are for the bug you already have; lectures are for the intuition you do not yet have, and two free series teach Part II better than any text for first exposure. The First Principles of Computer Vision series by Shree Nayar (Columbia) covers the imaging, features, and geometry of this part with exceptional clarity, and Cyrill Stachniss's Photogrammetry and Robotics lectures (Bonn) are the best free treatment of the calibration, SfM, and SLAM of Chapter 12 through Chapter 14. For the tools, the primary documentation is the reference, not a blog: the OpenCV tutorials for the modules of Section 17.1, and the COLMAP FAQ, which answers the overwhelming majority of reconstruction failures, including the disconnected-model problem from Section 17.2's practical example.

Staying current is a skill in itself, and the goal is sustainability, not completeness. Figure 17.4.1 lays out a low-effort weekly loop that catches the important developments without the firehose of reading everything.