The End Of Knowledge - Vault 5/64 - CVPR - 2021 - Computer Vision and the Global Goals

graph LR classDef goals fill:#f9d4d4, font-weight:bold, font-size:14px classDef gaps fill:#d4f9d4, font-weight:bold, font-size:14px classDef applications fill:#d4d4f9, font-weight:bold, font-size:14px classDef challenges fill:#f9f9d4, font-weight:bold, font-size:14px classDef solutions fill:#f9d4f9, font-weight:bold, font-size:14px A[Computer Vision and
the Global Goals] --> B[Vision aids sustainable development goals. 1] A --> C[Info scarcity, expert shortage hinder progress. 2] B --> D[Satellite imagery estimates population density. 3] B --> E[Automated disease diagnosis: malaria, TB, skin. 4-5] B --> F[Crop disease detection, pest quantification. 6] C --> G[Power, network, compute, memory constraints. 8] C --> H[Long-term maintenance, sustainability challenges. 9] C --> I[Local insight needed for problem selection. 10] C --> J[Domain expertise for metrics, labels, data. 11] C --> K[Permissions, trust-building not incentivized. 12] C --> L[Results questioning assumptions breed mistrust. 13] C --> M[Poor interventions can reinforce inequalities. 14] C --> N[Pilotitis: pilots fail to scale, endure. 15] A --> O[Local communities provide grounding, risk awareness. 16] O --> P[Data Science Africa: training, research, support. 17] O --> Q[End-to-end ownership by single teams succeeds. 18] O --> R[Country-level groups identify priorities, tech potential. 19] A --> S[CV4GC workshop engages global researchers. 20] S --> T[Development-relevant benchmarks, datasets needed. 21] S --> U[Low-resource model design for impact. 22] S --> V[Cross-disciplinary collaboration, local focus valuable. 23] S --> W[Real-world constraints inspire new CV methods. 24] A --> X[Impactful CV opportunities abound globally. 25] class A,B,C goals class D,E,F applications class G,H,I,J,K,L,M,N challenges class O,P,Q,R,S,T,U,V,W,X solutions

Resume:

1.- Computer vision can help address UN global goals for sustainable development in areas like health, climate, and others.

2.- Information gaps and shortage of experts in low-resource settings are two main areas where computer vision can help.

3.- Estimating population density using satellite imagery to count buildings is one application, useful for various purposes.

4.- Medical diagnosis, like detecting malaria parasites in blood smears, is limited by scarcity of lab technicians.

5.- Computer vision can help automate visual diagnostic tests for diseases like malaria, TB, intestinal parasites, and skin conditions.

6.- In agriculture, computer vision can help diagnose crop diseases and quantify disease vectors like insects on cassava plants.

7.- Many promising computer vision applications exist, but few are in real use in the developing world due to challenges.

8.- Technical constraints include limited power, network, compute and memory resources, and the need for end-to-end development.

9.- Long time scales, maintenance, and sustainability are also technical hurdles in deploying computer vision solutions.

10.- Identifying the right problem to solve requires a team with local understanding of nuances and subtleties.

11.- Defining precise metrics, labeling policies, and data collection requires domain knowledge from the local context.

12.- Organizational work like obtaining permissions and building trust with collaborators is crucial but not well incentivized.

13.- Lack of trust can manifest as bureaucracy or domain experts dismissing results that challenge their assumptions.

14.- Worst case, poorly designed interventions can cause harm by reinforcing inequalities or introducing fragilities.

15.- "Pilotitis" refers to the proliferation of small-scale pilots that don't scale or achieve lasting impact.

16.- Building local communities is one way to mitigate challenges by providing staying power, grounding, and understanding of risks.

17.- Data Science Africa is an example of a community providing training, research, networking, and mutual support.

18.- Single teams owning the end-to-end process from data collection to deployment is a common theme in successful projects.

19.- Country-level groups bringing together government, academia, and business can identify priority problems and technological opportunities.

20.- The Computer Vision for Global Challenges workshop at CVPR engaged a global group of researchers.

21.- Computer vision researchers can get involved by choosing benchmarks and data sets relevant to global development problems.

22.- Designing models for low-resource settings with limited compute and memory is another way to make research more relevant.

23.- Cross-disciplinary collaboration and community formation around local problems is valuable even for methodology-focused researchers.

24.- Real-world constraints can inspire new computer vision methodologies.

25.- Interesting problems with potential for computer vision solutions exist everywhere, not just in the developing world.

Knowledge Vault built byDavid Vivancos 2024